LEGITPIX

Research library

Ways to check a photo, video or recording — and what each misses.

Images, audio, video and their metadata. Covers major method families, public standards, representative research implementations, and documented commercial services. A broad research snapshot, not a claim to enumerate every paper, patent, private algorithm or current product.

  • 90 methods
  • 193 primary sources
  • Library updated 2 Nov 2026

Baseline research completed 6 October 2026; bounded primary-source maintenance reviews completed 12, 19 and 26 October and 2 November 2026. Source-specific dates show which references were revisited. Implementations were not installed or benchmarked, commercial terms were not negotiated, and no personal media was uploaded. Recheck the exact version and licenses before adoption.

Start with the claim

The same file can have valid provenance, match an older source, and still be presented with a false story. Read each result as evidence for a specific claim. The guide below is a synthesis; the catalog provides method-specific primary sources.

Exact identity

Are these the same bytes?

A cryptographic digest can support an exact-byte comparison to a specific receipt or trusted reference. Hashing begins when bytes are observed; it cannot establish their earlier history or the truth of the depicted event.

Source relationship

Is this a copy or derivative of something already indexed?

Perceptual hashes, local features and embeddings can retrieve candidate relationships despite some changes. Matching is threshold-dependent and bounded by the reference collection. Similarity alone does not establish ownership, a unique origin or semantic truth.

Embedded signal

Does this contain a recognized watermark?

A detector can support presence of a particular signal or recovery of a payload under its tested conditions. Validate the expected payload and any cryptographic authentication separately. A failed check may reflect unsupported marks, absent marking or damaged signals.

Signed provenance

Who signed which claims about this asset?

Check manifest binding, signature, identity and trust policy separately. A valid signature authenticates a signer's assertion and protected content relationship. The signer may be mistaken or dishonest, and a legitimate asset may have lost its manifest.

Capture attestation

What capture pipeline was attested?

Device- or app-backed capture can narrow the trusted path from capture to receipt. Its strength depends on hardware, app integrity, key custody and enrollment. Recording a staged scene or another screen remains a separate issue.

Time evidence

Did these bytes or this digest exist by a particular time?

A trusted timestamp or verifiable log can support existence by a time, subject to its trust model. A locally written timestamp is only a local assertion. Neither is automatically the time the depicted event happened.

Forensic inference

Do observed traces resemble manipulation or synthesis?

Statistical and learned detectors return evidence about a defined task and population. Scores need calibration, domain checks and error estimates. A clean result does not prove camera origin; an anomalous result does not by itself prove deception.

Contextual corroboration

Does the claimed story fit independent evidence?

Source history, other views, recorded events and expert analysis can corroborate context. A genuine file can carry a false caption, and synthetic media can accurately illustrate a real event. Content origin and factual claims require separate review.

A practical order for LegitPix

Recommended integration order, aligned with local receipt generation and verification. These are proposed stages; this reference library does not add features to the LegitPix app.

1 Preserve and identify

Keep the received file unchanged. Record SHA-256, byte length, content type, receipt schema/version and observation time locally. If normalized pixels or decoded audio are also hashed, label that derivative algorithm and decoding rules separately. A receipt should say exactly what was compared.

2 Inspect claims

Read metadata and C2PA manifests without treating their presence as validation. Separate absent, unsupported, malformed, cryptographically invalid, cryptographically valid but untrusted, and trusted under a named policy. Capture trust-list freshness and revocation limitations, especially offline.

3 Retrieve relationships

Add perceptual matching only against an authorized reference set. Show the matching reference, algorithm/version, distance and threshold, then require a source comparison before making stronger claims. Keep exact-byte equality and near-duplicate similarity separate.

4 Check specific marks

Use named watermark adapters with clear provider/model coverage, key or payload requirements and measured operating thresholds. Prefer local implementations when practical. Ask for permission before sending personal media to any external provider and disclose what leaves the device.

5 Analyze cautiously

Add passive image, audio and video detectors as qualified analysis after representative evaluation. Show the detector, model hash/version, input scope, score type, calibration population and limitations. Keep unsupported, not tested and inconclusive outcomes available.

6 Strengthen future capture

Evaluate signed receipts, independent timestamps and hardware-backed capture as separate extensions. Define exactly which trust boundary each extension strengthens. Do not relabel an existing imported photograph as sensor-attested.

Keep result labels specific.
  • Exact-byte match to receipt [receipt ID].
  • Signed provenance validates under [trust policy]; assertions shown separately.
  • Possible source match to [reference] at [algorithm/version/threshold].
  • Recognized [provider/method] watermark detected; payload/authentication status recorded.
  • Forensic indicator of [defined task]; score, scope and limitations available.
  • No supported evidence found, unsupported input, or inconclusive.
How to read the catalog

Each record combines the mechanism, supported claim, proof limits, implementation/access, offline feasibility, cost dependencies, code/model/data license notes, version/date, transformation behavior, error caveats, privacy, suggested LegitPix role, uncertainties and sources.

Evidence types are source categories, not quality scores. A standard defines rules. Research reports a particular study. Software repositories document an implementation. Vendor documentation describes a service. Publication, available code and production readiness are different milestones.

Offline-capable does not mean browser-ready. A downloadable model may need native libraries, large weights or a GPU. An external source URL opens only when you click it. Search and filters in this file run locally and send no media or queries anywhere.

Method counts are not an exhaustive census. Some records are families, others are named implementations or standards. Read the family and version fields before comparing entries.

Method catalog

Showing 90 of 90 methods

AI Detection· audio, video-audio-track, video

Audio anti-spoofing classifier (AASIST / ASVspoof baselines)

Learn spectro-temporal graph-attention representations to distinguish bona fide speech from synthetic/converted speech in defined tasks.

Show details
What it can support
A speech anti-spoofing score within the evaluated recording/channel/generator domain.
What it does not prove
Speaker identity, truthful words, consent, or universal synthetic-audio detection.
Maturity
ICASSP 2022 model; ASVspoof5 provides a 2024 baseline adaptation.
Implementation
Official AASIST/AASIST-L code and pretrained models; separate ASVspoof5 baseline.
Robustness to changes
Codecs, resampling, microphones, language and playback/re-recording create channel shifts.
Error caveats
A low challenge EER is not expected real-world error. Threshold, spoof prevalence and attack coverage change precision.
License and access
Original AASIST MIT; ASVspoof5 assets and speech datasets need separate license/consent review.
Works offline
Yes after model assets.
Cost and dependencies
Audio preprocessing and model inference; representative speech evaluation.
Privacy and rights
Speech contains personal/biometric information; local processing and limited retention.
Role for LegitPix
Optional speech-specific evidence, never a voice-authentication verdict.
Uncertainty
Benchmark task definitions and trained checkpoint must accompany results.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image

CNN-generated image classifier (CNNDetection baseline)

Supervised classifier learns synthesis artifacts, with blur/JPEG augmentation used to improve transfer across generation methods.

Show details
What it can support
A model score for resemblance to its training/evaluation classes.
What it does not prove
AI generation as fact, exact generator attribution, or real-world authenticity.
Maturity
Influential CVPR 2020 baseline; older generation coverage.
Implementation
Author PyTorch code, training/evaluation scripts and released weights.
Robustness to changes
Augmentation covers some blur/recompression; modern pipelines may differ substantially.
Error caveats
Author demo explicitly calls its output uncalibrated. Old benchmark performance cannot establish current-generator accuracy.
License and access
Repository CC BY-NC-SA 4.0; data/dependencies require separate review.
Works offline
Yes after assets are downloaded.
Cost and dependencies
Inference compute and representative evaluation data.
Privacy and rights
Do not accuse creators or make consequential decisions from a classifier output.
Role for LegitPix
Historical benchmark/control, not a production authenticity gate.
Uncertainty
No modern target-domain evaluation performed.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image

Diffusion reconstruction-error detection (DIRE)

Compare the input with a pretrained diffusion model's reconstruction, then classify the reconstruction-error representation.

Show details
What it can support
An input's compatibility with learned diffusion reconstruction/detection behavior.
What it does not prove
All AI generation, exact generator identity, untouched capture or event truth.
Maturity
ICCV 2023 paper and official research repository.
Implementation
DIRE code, diffusion reconstruction assets and detector checkpoints.
Robustness to changes
Reconstruction behavior depends on chosen diffusion model and input transformations.
Error caveats
Reported cross-diffusion transfer is conditional on tested generators/data; edited photos and new models need separate assessment.
License and access
No clear top-level license located in reviewed official repository; weights/data rights unverified.
Works offline
In principle after all assets; GPU-oriented and heavier than a simple classifier.
Cost and dependencies
Diffusion reconstruction latency/memory plus classifier.
Privacy and rights
Local processing possible; do not upload private images to a public demo without approval.
Role for LegitPix
Research/reference-only until rights and operational calibration are established.
Uncertainty
Public source availability is not permission for commercial redistribution.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image

Failure-focused detector training (PROBE)

Use difficult generative variations to refine detectors and improve transfer to unseen generation conditions.

Show details
What it can support
Research evidence that targeted variation coverage can improve a detector on evaluated benchmarks.
What it does not prove
Universal detection, robustness to all new generators, or calibrated authenticity probability.
Maturity
ICML 2026 method; early official repository with evaluation and checkpoint links.
Implementation
PROBE-AIGI-Detection; ResNet50 and DINOv2 detector checkpoints linked on ModelScope.
Robustness to changes
Research explores generation variation; ordinary distribution shifts remain a validation requirement.
Error caveats
More difficult examples can improve benchmark transfer without establishing real-world false-positive rates.
License and access
No top-level code license found in reviewed repository; checkpoint/data terms unverified.
Works offline
Potentially after assets; end-to-end training release completeness unverified.
Cost and dependencies
Inference backbone; training needs generator access and additional compute.
Privacy and rights
Training/evaluation data permissions and source integrity remain necessary.
Role for LegitPix
Current research watchlist and evaluation-design input.
Uncertainty
README has unchecked release TODOs alongside evaluation/checkpoint links; do not claim a complete training release.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image, video-frames, video

Foundation-model subspace adaptation (Effort)

Preserve principal pretrained feature components and adapt an orthogonal residual subspace to reduce detector overfitting.

Show details
What it can support
A trained model's anomaly/classification signal for the chosen face or general-image domain.
What it does not prove
Domain-independent accuracy, provenance, generator identity or an untouched image.
Maturity
ICML 2025 method with author code and checkpoints.
Implementation
Effort-AIGI-Detection; face-specific and general-image checkpoints use different training data/preprocessing.
Robustness to changes
Foundation knowledge may improve transfer; face crops and general images require the appropriate model.
Error caveats
Better benchmark transfer does not calibrate deployment scores. Wrong checkpoint/preprocessing can invalidate interpretation.
License and access
Top-level universal grant not located; bundled DeepfakeBench is CC BY-NC 4.0. Check each code subtree and checkpoint.
Works offline
Yes after checkpoint/dependency acquisition.
Cost and dependencies
CLIP-scale inference, preprocessing and evaluation.
Privacy and rights
Face processing is sensitive; avoid identity or intent conclusions.
Role for LegitPix
Research candidate for independently measured detector evidence.
Uncertainty
No weight license audit or runtime reproduction performed.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image

Frozen-foundation-feature detector (UniversalFakeDetect)

Use a frozen pretrained CLIP visual representation with a lightweight supervised detector to improve cross-generator transfer.

Show details
What it can support
Model-specific synthetic-image evidence within an evaluated domain.
What it does not prove
Universality, truthful camera origin, exact model attribution, or calibrated probability.
Maturity
CVPR 2023 public code and detector weights.
Implementation
UniversalFakeDetect PyTorch/CLIP; official repository moved to WisconsinAIVision.
Robustness to changes
Transfer evaluated across different generators; preprocessing and distribution shifts remain material.
Error caveats
A default 0.5 threshold is not a calibrated probability. Semantic/domain shortcuts and class imbalance can affect results.
License and access
Repository MIT; upstream weights and dataset rights must still be checked.
Works offline
Yes once detector and CLIP weights are present.
Cost and dependencies
Foundation-model inference and target-domain validation.
Privacy and rights
Prefer local inference and retain scores as tentative evidence.
Role for LegitPix
Candidate baseline for calibration experiments, with model/version recorded.
Uncertainty
Title's 'universal' describes a research aim, not a guarantee.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image, video-frames, video

Parameter-efficient cross-benchmark detection (GenD)

Adapt a pretrained visual encoder through LayerNorm tuning and normalized metric learning; evaluate cross-dataset deepfake transfer.

Show details
What it can support
Model-specific facial deepfake evidence under its preprocessing and training domain.
What it does not prove
Authenticity across all video types, calibration, or generator/source attribution.
Maturity
WACV 2026 paper, public implementation and pretrained model links.
Implementation
GenD PyTorch/Transformers, face preprocessing and model-specific backbone.
Robustness to changes
Must validate crops, compression, non-face video and modern generator changes.
Error caveats
Authors emphasize paired real/fake sources to reduce shortcut learning. Cross-benchmark AUROC is not a deployed probability.
License and access
Repository MIT; each backbone/checkpoint/data license must be reviewed separately.
Works offline
Yes after model assets; repository demonstrates local inference/UI.
Cost and dependencies
Backbone compute and face detection; GPU useful.
Privacy and rights
Limit face data retention and never infer personal identity from detector results.
Role for LegitPix
Current research candidate with explicit model and calibration profile.
Uncertainty
No checkpoint download or independent reproduction; backbone variants need separate review.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· video

Spatiotemporal facial deepfake detection (TALL)

Arrange sampled video frames into a thumbnail layout so a visual model can learn spatial/temporal inconsistencies.

Show details
What it can support
A model score for facial manipulation patterns in the sampled clip.
What it does not prove
Untampered complete video, truthful event context, intact audio or all-generator coverage.
Maturity
ICCV 2023 official code; 2024 extension announced.
Implementation
TALL4Deepfake PyTorch/Swin-based implementation and frame-sequence preparation.
Robustness to changes
Frame rate, temporal sampling, face visibility and recoding can affect the evidence.
Error caveats
A model can miss unsampled alterations; benign temporal artifacts can cause false alarms. Face-domain training is not general video authentication.
License and access
Official repository MIT; checkpoints/backbones/datasets separately reviewable.
Works offline
Yes with the necessary checkpoint/environment.
Cost and dependencies
Video decoding, face crops, frame sampling and model inference.
Privacy and rights
Limit face-frame retention and avoid identity conclusions.
Role for LegitPix
Research video-detector candidate, distinct from reference-based temporal matching.
Uncertainty
README still says TALL++ code will be made available; do not infer the extension is fully released.

Sources

Evidence type: research · Last reviewed 2026-10-06

AI Detection· image

TruFor manipulation detection and localization

Fuse RGB and learned noise-sensitive features to produce manipulation localization, an image-level score and a reliability map.

Show details
What it can support
Model-estimated local manipulation anomalies and associated uncertainty.
What it does not prove
That every clean-looking image is genuine, which editor was used, or what happened in the scene.
Maturity
CVPR 2023 framework; test and training code publicly available.
Implementation
Author PyTorch implementation; Docker inference documented.
Robustness to changes
Evaluated across conventional and learning-based manipulation types; operational transformations still need testing.
Error caveats
Localization/reliability maps are model outputs, not calibrated proof. Domain shift, benign processing and small edits can cause errors.
License and access
Informational/nonprofit-only author license; no unrestricted commercial deployment inferred.
Works offline
Yes after weights and environment are acquired.
Cost and dependencies
Typically GPU-friendly inference and substantial model dependencies.
Privacy and rights
Keep original and derived maps private; maps can be reputationally consequential if overstated.
Role for LegitPix
Opt-in research anomaly maps with uncertainty and expert review.
Uncertainty
Noiseprint++ training release remained a TODO in reviewed README; no runtime test.

Sources

Evidence type: research · Last reviewed 2026-10-06

Audio Matching· audio, video-audio-track, video

Chromaprint compact audio fingerprints

Generate compact fingerprints intended for near-identical recordings, duplicates and long-stream monitoring.

Show details
What it can support
Recording similarity against a compatible reference index.
What it does not prove
Speaker identity, truthful speech, ownership, or absence of audio synthesis.
Maturity
Established library and fpcalc utility.
Implementation
acoustid/chromaprint; FFmpeg commonly supplies decoding; FFT backend configurable.
Robustness to changes
Optimized for near-identical audio; not a general-purpose noisy short-clip recognizer.
Error caveats
Library explicitly trades precision/robustness for search speed; silence, short clips and major alterations need special handling.
License and access
Own code MIT, but LICENSE.md says project as a whole LGPL-2.1 because included FFmpeg code; FFTW can impose GPL obligations.
Works offline
Yes for fingerprint computation and a separately implemented/local matcher.
Cost and dependencies
Local compute; AcoustID service/index is optional and separate.
Privacy and rights
Audio fingerprints still permit linkage to recordings.
Role for LegitPix
Optional recording-match evidence; do not confuse with an audio watermark detector.
Uncertainty
Audit exact source version and compiled dependency licenses before shipping.

Sources

Evidence type: software · Last reviewed 2026-10-06

Audio Matching· audio, video-audio-track, video

Landmark-based audio fingerprinting (audfprint)

Index spectral landmarks and find temporally consistent matches for noisy excerpts.

Show details
What it can support
Candidate recording/segment reuse with an estimated offset.
What it does not prove
What was said is true, who spoke, or whether the source is synthetic.
Maturity
Established author-maintained research implementation; current support cadence not established.
Implementation
Dan Ellis audfprint Python package; FFmpeg decoding.
Robustness to changes
Designed for excerpts and noise; timing/encoding/channel changes still require evaluation.
Error caveats
Shared music, jingles and low-information sound can produce ambiguous matches. Distinguish hash collisions from sustained aligned evidence.
License and access
MIT code; FFmpeg and Python dependencies separately licensed.
Works offline
Yes; local database creation and queries.
Cost and dependencies
CPU, reference database size and landmark density.
Privacy and rights
Audio indexing may expose speech/music membership; keep authorized local references.
Role for LegitPix
Audio-segment relationship matcher with offsets and matched duration.
Uncertainty
No production scalability or target-domain accuracy claim verified.

Sources

Evidence type: software · Last reviewed 2026-10-06

Audio Matching· audio, video-audio-track, video

Panako time/pitch-tolerant audio matching

Audio fingerprinting research system with algorithms designed to recognize time-scale and pitch-modified recordings.

Show details
What it can support
A candidate recording relationship under wider timing/pitch variation.
What it does not prove
Speaker authenticity, musical ownership, or truthful content.
Maturity
Research platform with public code and documented evaluation scripts.
Implementation
Java/Gradle, FFmpeg, LMDB and Gaborator-related dependencies.
Robustness to changes
Explicit focus on speed, time-stretch and pitch variation.
Error caveats
Claims apply to tested material/settings; do not transfer music retrieval performance to all speech/noise.
License and access
AGPL source; dependency licenses vary. README raises patent considerations; present enforceability not assessed.
Works offline
Yes after installation; platform-specific dependencies matter.
Cost and dependencies
Compute and local index; potentially significant integration/legal review.
Privacy and rights
Use permitted reference audio; database access can reveal listening or recording relationships.
Role for LegitPix
Specialized research candidate when simple audio matching is inadequate.
Uncertainty
README contains stale platform statements alongside newer updates; verify current build before adoption.

Sources

Evidence type: software · Last reviewed 2026-10-06

Capture· image, video

Controlled capture and signing: Truepic

Controlled capture with product-specific assurance: native Vision uses device-key signing and device-integrity checks; browser and embedded-WebView approaches use server-side signing. Separate Lens server products can sign existing files.

Show details
What it can support
Evidence about an explicitly identified controlled capture or signing workflow.
What it does not prove
Every Truepic signature originated at capture, or that a scene/caption is factual. Server signing is a separate workflow.
Maturity
Commercial capture/service ecosystem
Implementation
Truepic mobile/browser capture integrations and Lens server CLI/API; verify exact product scope.
Robustness to changes
C2PA hard-binding rules apply; capture checks and downstream edit history remain distinct.
Error caveats
Browser/embedded approaches do not provide native device attestation or root/jailbreak checks. Vendor assurance claims require exact product, platform and threat model; independent universal error rates unavailable.
License and access
Proprietary contracts; Lens CLI/display library treated as confidential in current MSA.
Works offline
Native Vision supports offline capture. Web/embedded approaches provide pending-upload retry rather than a full offline workflow. Lens server enrollment requires connectivity; timestamping may require it.
Cost and dependencies
Enterprise integration, CA/service access and negotiated price; no free general detector confirmed.
Privacy and rights
Capture may involve location/device attestation; server Lens MSA describes usage telemetry. Minimize and review contract.
Role for LegitPix
Optional capture-origin adapter; label capture evidence separately from post-hoc signing.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-12

Capture· image

Hardware-backed capture signing: Leica M11-P example

Camera signs captured images using protected hardware and provisioned certificates when Content Credentials is enabled.

Show details
What it can support
A valid credential attributes captured-file evidence to the camera’s signing implementation.
What it does not prove
The scene was unstaged, correctly captioned, or not a photograph of another image/display.
Maturity
Commercial hardware example
Implementation
Leica M11-P integrated signing; independent C2PA viewers can inspect supported outputs.
Robustness to changes
Direct signatures bind original output; derivatives need preserved provenance or new compliant signing.
Error caveats
Validate actual certificate/firmware and signed fields; camera claims do not make arbitrary EXIF true.
License and access
Proprietary firmware/hardware; open C2PA format.
Works offline
Camera capture signing and local signature verification possible; trust/status freshness remains separate.
Cost and dependencies
Specific camera hardware and lifecycle support; current purchase price not evaluated.
Privacy and rights
Device/creator metadata can identify a photographer; inspect what is published.
Role for LegitPix
Higher-assurance capture-origin evidence when present; never prerequisite for a legitimate image.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor/standard · Last reviewed 2026-10-06

Capture· image, video

Sony Camera Authenticity Solution

Hardware signing combines C2PA with proprietary depth/recapture analysis, modification analysis and server-time evidence.

Show details
What it can support
Valid camera-signing evidence plus separately reported vendor analysis under supported capture conditions.
What it does not prove
Scene/event truth or harmless intent. A real 3D object can still be staged or misleading.
Maturity
Commercial restricted professional workflow; public-reporting/display tools include Beta components.
Implementation
2026.2 service; compatible still/video cameras; Sony Self-Checker and validation site. Image Validation API supports still images only; pre-check SDK supports stills and video. Camera Verify, Stamping Tool and SignatureViewer.js remain Beta.
Robustness to changes
Compatible editing can preserve history; specific video-format/bitrate restrictions apply.
Error caveats
Do not equate generic C2PA validation with Sony’s proprietary 3D or modification results; independent error rates not established.
License and access
Proprietary hardware/software/service; C2PA format open.
Works offline
Basic local signature/pre-check possible; advanced Sony validation is subscription-dependent.
Cost and dependencies
Paid digital-signature upgrade and validation service; default certificate is free but trust-list status must be checked.
Privacy and rights
Advanced verification may upload media/depth metadata; authorization required.
Role for LegitPix
Import granular external evidence if available; avoid exclusive vendor dependency.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-12

Exact Integrity· image, video, audio

Canonical decoded-content and per-frame hashing

Hash decoded pixels or audio samples under an explicitly fixed decoding/canonicalization profile; FFmpeg supplies hash, framehash and streamhash muxers.

Show details
What it can support
Equality of selected decoded data under the recorded profile, including frame-level comparison.
What it does not prove
Identical container bytes, all metadata, temporal context, or authenticity.
Maturity
Mature diagnostic tools; a LegitPix canonicalization profile still needs specification.
Implementation
FFmpeg; pin decoder/build, streams, pixel/sample format, dimensions, color handling and timing policy.
Robustness to changes
Container-only changes may leave decoded content equal; resizing, color conversion or lossy recoding usually will not.
Error caveats
Different decoders or color/sample conversions may produce different bytes. Equal sampled frames do not cover omitted frames.
License and access
FFmpeg usually LGPL-2.1-or-later; enabled components can make a build GPL or otherwise restricted.
Works offline
Yes.
Cost and dependencies
Decoding cost; codec support and build variability.
Privacy and rights
Local processing preferred; preserve raw original separately.
Role for LegitPix
Optional diagnostic relationship below whole-file integrity.
Uncertainty
No universal decoded-media canonical form is specified here; not runtime-tested.

Sources

Evidence type: software · Last reviewed 2026-10-06

Exact Integrity· image, video, audio, any-file

Cryptographic whole-file hashing (SHA-256)

Hash the exact received bytes and compare with a trusted recorded digest.

Show details
What it can support
Extremely strong byte-identity/change evidence under the hash's security assumptions.
What it does not prove
Capture origin, author, time, consent, truth, or whether the reference was already manipulated.
Maturity
Mature standard and ubiquitous implementations.
Implementation
Standard cryptographic libraries; retain algorithm and digest with receipt.
Robustness to changes
Any byte change normally changes the digest, including harmless metadata edits.
Error caveats
A matching hash without an authenticated reference says only that two inputs match. Hashing alone is not signing.
License and access
Algorithm standard; chosen implementation has its own license.
Works offline
Yes; no media upload required.
Cost and dependencies
Negligible CPU; trust in the reference record is separate.
Privacy and rights
Hashes can link records and permit known-file membership tests; do not treat them as anonymized.
Role for LegitPix
Core byte-consistency receipt; bind claims to the original file digest.
Uncertainty
Does not establish truth before the first trusted observation.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Hosted AI Detection· image, video, audio

Commercial AI-generated and face-deepfake detection (Hive)

Hosted visual API combines generated-image/video classification with face-deepfake classification; a separate audio model analyzes synthetic audio.

Show details
What it can support
Provider-scoped model scores and tentative generator-family labels.
What it does not prove
A trusted generator attribution, video/audio integrity, absence of manipulation or rights ownership.
Maturity
Documented commercial APIs; provider claims only in this catalog.
Implementation
Authenticated API; record visual, face and audio outputs separately. Visual video results include sampled-frame decisions.
Robustness to changes
Frame sampling, compression, face visibility and audio channel changes affect applicability.
Error caveats
Vendor recommends use-case threshold refinement. Maximum face score and per-frame aggregation can amplify errors across many faces/frames.
License and access
Proprietary models/service terms; not an open-source detector.
Works offline
Hosted service; no public offline weights verified. Private deployment terms unconfirmed.
Cost and dependencies
Usage-based service; image, video-frame and audio processing have different units. Current contract and total workload cost not verified.
Privacy and rights
FAQ states default customer-data retention of 14 days, adjustable on request; zero retention must be established for the actual project.
Role for LegitPix
Optional external detector evidence; distinguish neural scores from any separately returned C2PA metadata.
Uncertainty
No customer account, private deployment, model-version guarantee or independent performance test performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Hosted AI Detection· image, video, audio

Commercial image/video, deepfake and audio detection (Sightengine)

Hosted models offer generated-image/video, face-deepfake, music and speech analysis, with global and some generator-specific scores.

Show details
What it can support
A provider-produced detector observation for the selected model and input.
What it does not prove
True provenance, exact generating system, genuine voice identity or trustworthy event content.
Maturity
Operating visual APIs; speech model documented as beta and gated.
Implementation
Authenticated image/video/audio APIs; preserve modality-specific scores and generator-label schema. October 2026 expands named image/video classes previously returned as other.
Robustness to changes
Frame selection, image rescaling and audio channel changes need separate validation.
Error caveats
Performance and generator labels remain provider claims; scores across modalities/models are not automatically comparable or calibrated.
License and access
Proprietary subscription/model terms; checkpoint licenses not available.
Works offline
Hosted API; no public offline model distribution or private deployment verified.
Cost and dependencies
Plans and operation/model-based billing; speech docs restrict access to enterprise users/partners, despite broader paid-plan FAQ wording.
Privacy and rights
Uploads or fetched media URLs expose content to the service. Confirm subscription retention/subprocessors; privacy policy allows agreement-dependent retention.
Role for LegitPix
Optional gated vendor evidence after scope, privacy and threshold evaluation.
Uncertainty
April 2026 guide still describes speech as in development; current June changelog/docs used. No account, media upload or contractual audit.

Sources

Evidence type: vendor · Last reviewed 2026-10-12

Hosted AI Detection· image, audio, video

Commercial multimodal deepfake detection (Reality Defender RealAPI)

Proprietary models return ensemble manipulation scores and supporting indicators through an authenticated upload/API workflow.

Show details
What it can support
A provider-generated analysis result for the submitted media and model version.
What it does not prove
Authenticity, identity, event truth or a calibrated real-world probability merely because the provider calls its score a probability.
Maturity
Operating commercial API and enterprise products; independent performance not established here.
Implementation
REST/SDK file upload then polling or webhook; account/API key required.
Robustness to changes
Compression, media quality and unseen generators require target-specific testing.
Error caveats
Treat claimed performance as vendor evidence. Preserve modality/model details and independently calibrate thresholds.
License and access
Proprietary service/models; SDK availability does not license model weights.
Works offline
Hosted API requires network; enterprise page advertises on-premises, private-cloud, containerized and air-gapped laptop deployments, subject to contract.
Cost and dependencies
Free tier lists image/audio; video is included in paid tiers. Enterprise pricing/deployment quoted separately; billing details not audited.
Privacy and rights
Hosted uploads leave the device; organization admins can view users' uploads. Verify retention, locality and contractual processing before use.
Role for LegitPix
Optional commercial adapter after data-sharing approval and validation; never collapse its result into verified provenance.
Uncertainty
Older FAQ access language conflicts with current self-serve product page; current product page used. No account, upload, quote or private deployment tested.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Metadata· image, video

EXIF/XMP/IPTC metadata and Digital Source Type

Structured fields describe camera, creator, location, software, rights and declared source type; IPTC adds AI system/prompt fields.

Show details
What it can support
The file contains these declarations; stronger attribution requires authenticated inclusion in a signed assertion.
What it does not prove
Values are true, the named camera produced the pixels, absence of AI, or complete editing history.
Maturity
Widely deployed descriptive metadata
Implementation
Common metadata parsers/editors; IPTC 2025.1 current published standard.
Robustness to changes
Platforms/editors may alter or discard fields; changing covered metadata can invalidate C2PA.
Error caveats
Treat unsigned source-type and software fields as claims, not positive or negative AI detection.
License and access
IPTC specification CC BY 4.0; format/tool terms separate.
Works offline
Yes; conventional parsing.
Cost and dependencies
Open tools available; no mandatory hosted service.
Privacy and rights
GPS, names and prompts may be sensitive; do not expose all metadata by default.
Role for LegitPix
Context/explanation layer labelled “declared metadata,” separate from signed evidence.
Uncertainty
This record chiefly covers photo/video metadata standards. Audio container-specific tags and applications need separate treatment.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Multi Level Identifier· text, image, audio, video, any-file

ISCC multi-component content identifier (ISO 24138)

Build a typed composite identifier from metadata, perceptual/content, data-similarity and exact-instance components.

Show details
What it can support
Component-specific exact or similarity relationships for a digital asset.
What it does not prove
Ownership, registered authorship, trusted capture or truth of content.
Maturity
ISO standard with open reference implementation.
Implementation
iscc-core for core algorithms; higher-level extraction is delegated to iscc-sdk.
Robustness to changes
Similarity preservation depends on selected unit; instance checksum remains exact.
Error caveats
Do not interpret a soft component as a cryptographic commitment. Registration and trustworthy declarations are separate.
License and access
iscc-core Apache-2.0; full ISO text and SDK dependencies have separate terms.
Works offline
Core generation yes; format extraction/dependencies and any registry service are separate.
Cost and dependencies
Local computation; optional search/index/registry infrastructure.
Privacy and rights
Identifiers enable cross-system linkage and membership tests.
Role for LegitPix
Interoperable optional asset identifier alongside byte receipts and explicit evidence states.
Uncertainty
Semantic-Code is reserved, not algorithmically standardized by the 2024 edition; experimental extensions must be labeled.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Multimodal Forensics· audio, video

Audio-visual correspondence and fusion (AVFF representative)

Learn correspondences between sound and visual motion, then combine modalities for video deepfake classification.

Show details
What it can support
Model-estimated audiovisual inconsistency or synthetic-content evidence.
What it does not prove
That synchronized media is genuine, that a dubbed video is deceptive, or that audio/video share a trusted source.
Maturity
CVPR 2024 research; official distributable implementation not verified.
Implementation
AVFF paper describes self-supervised correspondence learning followed by supervised classification.
Robustness to changes
Dubbing, translation, latency, edits and audiovisual offsets can change consistency without deception.
Error caveats
Matching lip motion can be synthesized; off-screen speech or occlusion can invalidate assumptions. Dataset-specific scores cannot be generalized.
License and access
Publication access is not a software/model license; third-party reimplementations differ.
Works offline
Conceptually possible; official complete code/weights not established in this review.
Cost and dependencies
Two-modality models, synchronization and sufficiently visible speaking subjects.
Privacy and rights
Combined face and voice analysis is sensitive; do not infer identity.
Role for LegitPix
Reference/research taxonomy for corroborating modalities, not a deployable default.
Uncertainty
No official implementation/license or operational calibration verified.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· audio, video-audio-track, video

Audio recording and edit forensics

Compare waveform, spectrogram, encoding structure, bit depth, channel behavior and recording-device exemplars against a specific production claim.

Show details
What it can support
Consistency or inconsistency with an alleged recording process, including some edits or pause/restart events.
What it does not prove
Complete absence of editing, speaker identity, intent or truth of recorded statements.
Maturity
Established specialist methodology
Implementation
Expert workflow combining file and signal examination; preserve the original encoded file.
Robustness to changes
Re-encoding can discard format evidence and change traces; decoded audio can retain some signal-level evidence.
Error caveats
Ordinary device operations can resemble edits; conclusions require validated methods and controlled exemplars.
License and access
Guidance is not a software license. Tool and reference-data rights must be checked separately.
Works offline
Local examination is possible; exemplars, specialist tools or reference data may be needed.
Cost and dependencies
Analyst time, tools and reference data; no prices verified.
Privacy and rights
Recordings and exemplars can contain private conversations and third-party voices.
Role for LegitPix
Explain specialist evidence and support escalation beyond automated scores; not a first-release verdict engine.
Uncertainty
No forensic examination or tool benchmark performed. Guidance is not product certification or endorsement.

Sources

Evidence type: professional guidance · Last reviewed 2026-10-06

Passive Forensics· image

CFA/demosaicing consistency analysis

Estimate periodic color-interpolation traces, such as Bayer mosaic alignment, and flag locally inconsistent patterns.

Show details
What it can support
An image region is inconsistent with a particular assumed sampling/demosaicing pattern.
What it does not prove
Specific device identity, universal editing detection or scene truth.
Maturity
Established niche forensic method; reviewed IPOL implementation.
Implementation
Intermediate Values Algorithm for mosaic analysis.
Robustness to changes
Scaling, rotation, compression and modern multi-frame processing can weaken or alter traces.
Error caveats
Bayer assumptions do not cover every sensor or computational-imaging pipeline. Pattern agreement cannot rule out editing.
License and access
IPOL code metadata AGPL-3.0-or-later; version 1.0.
Works offline
Yes; source distribution available.
Cost and dependencies
CPU and suitable sufficiently preserved image data.
Privacy and rights
Local analysis preserves confidentiality; retain raw original if available.
Role for LegitPix
Specialist capture-pipeline consistency evidence.
Uncertainty
Applicability should be checked before interpreting any heatmap.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· audio, video-audio-track, video

Electric network frequency evidence in recordings

Analyze time-varying residual mains hum and compare it with suitable historical grid references when a usable signal exists.

Show details
What it can support
Qualified evidence about timing or continuity under validated extraction and comparison conditions.
What it does not prove
Guaranteed recording date, precise location, authenticity of speech or absence of editing.
Maturity
Established specialist methodology
Implementation
Specialist ENF extraction and database comparison; no turnkey implementation assessed.
Robustness to changes
Noise, duplication, processing and recording conditions can obscure or confound the signal.
Error caveats
Hum may be absent. Database coverage, cyclic patterns and multiple signals complicate matching; consider no-reference outcomes.
License and access
Guidance is not a software license. Tool and reference-data rights must be checked separately.
Works offline
Local examination is possible; exemplars, specialist tools or reference data may be needed.
Cost and dependencies
Analyst time, tools and reference data; no prices verified.
Privacy and rights
Inferred recording time or grid association can reveal context; protect reference and case data.
Role for LegitPix
Explain specialist evidence and support escalation beyond automated scores; not a first-release verdict engine.
Uncertainty
No forensic examination or tool benchmark performed. Guidance is not product certification or endorsement.

Sources

Evidence type: professional guidance · Last reviewed 2026-10-06

Passive Forensics· image

Error Level Analysis (ELA) visualization

Compare an image with a recompressed copy and visualize differences.

Show details
What it can support
Only that regions respond differently to that chosen recompression procedure.
What it does not prove
Forgery, AI generation, edit count or authentic untouched regions.
Maturity
Accessible inspection aid; weak standalone evidential value.
Implementation
Forensically browser-local ELA; author explicitly warns results can mislead.
Robustness to changes
JPEG quality, edges, textures, previous exports and re-encoding strongly change the visualization.
Error caveats
Brightness/heatmap contrast is not tampering probability. Natural high-detail regions can look suspicious; manipulations can look uniform.
License and access
Free use documented; redistribution license for the complete app not verified.
Works offline
Forensically says yes in modern browsers after assets are cached; images remain local.
Cost and dependencies
Free web tool and local browser compute.
Privacy and rights
Provider documents no image uploads; verify behavior of any deployment or fork.
Role for LegitPix
Expert inspection overlay only; never drive a green/red authenticity badge.
Uncertainty
No app runtime/network inspection performed; offline/privacy behavior is author-reported.

Sources

Evidence type: software · Last reviewed 2026-10-06

Passive Forensics· image, audio, video

EXIF/XMP/IPTC/container metadata and thumbnail consistency

Read metadata, embedded previews, software tags, timestamps and file structure; compare claims for consistency.

Show details
What it can support
What the file presently declares, plus inconsistencies that merit investigation.
What it does not prove
That a timestamp, camera identifier, location, author or software field is truthful.
Maturity
Mature parsing tools and forensic workflow.
Implementation
Read-only ExifTool extraction; record raw and interpreted values, duplicate fields and parser version.
Robustness to changes
Re-export, messaging platforms and privacy tools may remove or rewrite metadata and thumbnails.
Error caveats
Missing tags are ordinary; editing-software tags do not prove deceptive editing. Metadata can be altered independently of pixels.
License and access
ExifTool under Perl's terms: Artistic License or GPL; dependencies separate.
Works offline
Yes.
Cost and dependencies
Low local compute; format support/version matters.
Privacy and rights
GPS, names and device serials may be sensitive; avoid external extraction uploads.
Role for LegitPix
Descriptive evidence with clear unverified-claim status; signed provenance evaluated separately.
Uncertainty
Do not conflate untrusted metadata with signed content credentials.

Sources

Evidence type: software · Last reviewed 2026-10-06

Passive Forensics· image

Internal copy-move detection

Search within an image for duplicated patches/descriptors and geometrically consistent displacement fields.

Show details
What it can support
Candidate repeated regions that may result from cloning or object duplication.
What it does not prove
Copying direction, deceptive intent, external splicing, or that naturally repetitive content was edited.
Maturity
Established method family with reviewed reference implementations.
Implementation
IPOL implementation of dense-field copy-move detection by Ehret, discussing Cozzolino et al.
Robustness to changes
Methods can cover translations and some rotated/scaled copies; tolerance depends on descriptors/settings.
Error caveats
Repeated windows, foliage, textures or generated repeated structures can trigger candidates. Heavy changes and small regions can be missed.
License and access
IPOL source metadata lists GPL-3.0-or-later; code v1.0.
Works offline
Yes; source archive available.
Cost and dependencies
Memory/compute rises with image size and dense descriptor processing.
Privacy and rights
Local analysis; retain input and output masks without modifying evidence.
Role for LegitPix
Region-level evidence with side-by-side correspondences, not an automatic forgery verdict.
Uncertainty
Paper page intermittently failed; indexed primary record verified title, method and software license.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· image

JPEG grid and compression-history analysis

Recover local JPEG block-grid origins or DCT-statistical traces; compare inconsistent grids/compression evidence.

Show details
What it can support
Potential processing-history or region inconsistency compatible with compositing.
What it does not prove
Malicious tampering, how many legitimate edits occurred, or that consistent compression means authenticity.
Maturity
Reviewed IPOL reference algorithms with source and model-based false-alarm controls.
Implementation
Local JPEG Grid Detector and ZERO; other double-JPEG tools need their own validation.
Robustness to changes
Recompression, crops and resizing can create, erase or shift traces.
Error caveats
Ordinary editing/export can cause multiple-compression evidence. Model-based NFA is not a posterior authenticity probability.
License and access
Both cited IPOL code records AGPL-3.0-or-later; article rights separate.
Works offline
Yes from source.
Cost and dependencies
Local CPU and knowledgeable interpretation.
Privacy and rights
Analyze a working copy and retain exact file bytes.
Role for LegitPix
Specialist supporting signal; log detected grids, masks and parameters.
Uncertainty
Not every codec or camera pipeline is covered; no benchmark run.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· image, video

Lighting, shadows and scene-geometry consistency

Compare physically modeled illumination or geometry across objects/frames using explicit scene assumptions.

Show details
What it can support
Inconsistency with a specified lighting/geometric model; a reason for further investigation.
What it does not prove
Manipulation from visual oddness alone, or authenticity from plausible geometry.
Maturity
Established specialist method family; not a universal one-click detector.
Implementation
Johnson/Farid illumination research provides point-source and more complex lighting models.
Robustness to changes
Unknown light sources, reflections, perspective, lens distortion and processing change interpretation.
Error caveats
Incorrect physical assumptions can create false alarms; plausible synthetic scenes may satisfy simple models.
License and access
Papers do not provide a blanket software license; choose/verify any implementation separately.
Works offline
Yes if suitable images/calibration and analysis tools are available.
Cost and dependencies
Expert time, sufficient scene detail and potentially additional views.
Privacy and rights
Scene reconstruction can expose location or private context.
Role for LegitPix
Expert corroboration, with assumptions, alternatives and uncertainty disclosed.
Uncertainty
No ready maintained implementation or measured operating threshold established.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· image

Noise-level inconsistency (NoiseSniffer)

Use a background stochastic noise model to localize anomalies and report a number-of-false-alarms statistic.

Show details
What it can support
Anomalous local noise relative to specified model assumptions.
What it does not prove
That an anomaly is malicious editing, that noise-consistent content is authentic, or the source device.
Maturity
IWBF 2021 author implementation with reproducible output masks/statistics.
Implementation
Python NoiseSniffer and scientific-library dependencies.
Robustness to changes
Denoising, tone processing, recompression and heterogeneous image regions influence noise estimates.
Error caveats
A model-based false-alarm statistic is not the probability an image is fake; assumptions may fail on computational photography.
License and access
Author repository GPL-3.0; inspect exact license/dependencies before redistribution.
Works offline
Yes.
Cost and dependencies
CPU/memory and image resolution.
Privacy and rights
Local processing available; preserve original and analysis parameters.
Role for LegitPix
Complementary statistical anomaly signal, separated from learned detectors.
Uncertainty
No independent deployment validation; later IPOL refinement exists but full page was unavailable during this pass.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· image

Noiseprint learned camera-model traces

A CNN extracts noise-like processing traces associated with camera models and supports blind localization of inconsistent regions.

Show details
What it can support
Processing-trace inconsistency or similarity under the trained model.
What it does not prove
A specific physical camera, full provenance, or an untouched/truthful scene.
Maturity
Public 2019 research code; legacy dependency stack.
Implementation
Author TensorFlow implementation and models; blind heatmap/extraction tools.
Robustness to changes
Camera processing, recompression and resizing can alter residual patterns.
Error caveats
Benign local processing or unusual capture pipelines can appear anomalous; weak maps can be inconclusive.
License and access
Author license restricts use/reproduction/modification to informational and nonprofit purposes.
Works offline
Yes once assets/dependencies are available.
Cost and dependencies
Inference compute; older TensorFlow environment may need isolation.
Privacy and rights
Local analysis preferred; learned traces can support camera-family linkage.
Role for LegitPix
Research-only forensic layer pending rights and validation.
Uncertainty
Do not label this PRNU device identification; maintenance and production compatibility unverified.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· image, video-frames, video

Sensor-pattern / PRNU source-camera comparison

Estimate photo-response non-uniformity from reference captures and test correlation against a questioned image or region.

Show details
What it can support
Conditional support for a shared sensor source or locally inconsistent sensor trace.
What it does not prove
Scene truth, original capture time, photographer, or universal unique-device attribution.
Maturity
Established forensic method requiring specialist validation.
Implementation
Polimi ISPL prnu-python, referencing Binghamton's MATLAB implementation.
Robustness to changes
Resizing, stabilization, denoising, compression and computational photography affect extractability/alignment.
Error caveats
Modern processing can introduce non-unique artifacts and false matches. Non-detection is not proof of synthetic content.
License and access
prnu-python MIT; other implementations/data separately licensed.
Works offline
Yes; reference captures and careful alignment are normally required.
Cost and dependencies
Reference acquisition, compute, expertise and device-specific validation.
Privacy and rights
Device linkage can reveal private source relationships; restrict reference databases.
Role for LegitPix
Specialist source-consistency evidence with controls, not default authenticity scoring.
Uncertainty
Unique-sensor conclusions cannot be transferred uncritically across modern devices and modes.

Sources

Evidence type: research · Last reviewed 2026-10-06

Passive Forensics· video, audio

Video timeline, frame correlation and container/codec forensics

Inspect timing, frame relationships, container structure, GOP/compression evidence and source-device exemplars; assess audio separately.

Show details
What it can support
Consistency or inconsistency with the claimed capture/processing history.
What it does not prove
Truth of the filmed event, lack of staging, or authenticity of an unexamined audio stream.
Maturity
Established forensic examination practice.
Implementation
SWGDE workflow plus validated file/frame inspection tools; exemplar-device comparisons where available.
Robustness to changes
Transcoding, frame-rate conversion, stabilization and legitimate editing can explain apparent discontinuities.
Error caveats
Missing/duplicated frames need contextual explanation; heavily processed video can be inconclusive. Avoid absolute certainty.
License and access
Guidance is not software; SWGDE redistribution conditions apply to its documents.
Works offline
Yes for file examination.
Cost and dependencies
Analyst expertise, original-quality media and source exemplars.
Privacy and rights
Preserve original and audit trail; restrict video/audio evidence access.
Role for LegitPix
Specialist multi-signal report; never substitute temporal smoothness for authenticity.
Uncertainty
SWGDE lists method families; tool-specific accuracy still requires validation.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Perceptual Matching· image

Classical perceptual image hashes: aHash, dHash, pHash, wHash, color and crop-resistant hashes

Compact descriptors summarize luminance averages, gradients, DCT/wavelet structure, color, or segmented image regions; compare using algorithm-appropriate distances.

Show details
What it can support
Candidate near-duplicate or region relationships to an indexed reference.
What it does not prove
Cryptographic integrity, common author, originality, or image truth.
Maturity
Established lightweight baselines.
Implementation
JohannesBuchner ImageHash, Pillow/NumPy/SciPy/PyWavelets dependencies as applicable.
Robustness to changes
Some tolerate moderate resizing/recompression; rotation, crops, overlays and color changes affect variants differently.
Error caveats
Small hashes collide on simple/repetitive images. Hamming thresholds are dataset-specific; color similarity is especially weak identity evidence.
License and access
ImageHash BSD-2-Clause; dependencies separately licensed.
Works offline
Yes; CPU-friendly.
Cost and dependencies
Local compute and reference index.
Privacy and rights
A hash can reveal database membership; index only media you may retain.
Role for LegitPix
Low-cost candidate retrieval followed by stronger comparison.
Uncertainty
Variants are not interchangeable; no universal threshold or verified robustness ranking.

Sources

Evidence type: software · Last reviewed 2026-10-06

Perceptual Matching· image, video-frames, video

Local keypoints plus geometric verification (SIFT/ORB)

Match local descriptors, then test geometrically consistent correspondences using robust estimation such as RANSAC.

Show details
What it can support
A shared visual region and an estimated geometric relationship.
What it does not prove
Whole-image identity, copying direction, authorship, or scene authenticity.
Maturity
Mature computer-vision building blocks.
Implementation
OpenCV feature2d and homography tools; SIFT or ORB with appropriate descriptor distance.
Robustness to changes
Can accommodate scale, rotation, crop and perspective within model assumptions.
Error caveats
Repeated textures can yield false correspondences; low-texture or highly changed regions can fail. Planar homography is not universal 3D geometry.
License and access
OpenCV 4.5+ Apache-2.0; inspect bundled dependencies.
Works offline
Yes; CPU implementations available.
Cost and dependencies
Feature extraction and candidate-pair verification.
Privacy and rights
Avoid repurposing descriptors for identity tracking without an appropriate basis.
Role for LegitPix
Confirm candidate region reuse with match visualizations and inlier counts.
Uncertainty
Tutorial thresholds are examples, not validated operating points.

Sources

Evidence type: software · Last reviewed 2026-10-06

Perceptual Matching· image

PDQ image similarity hashing

Meta's PDQ produces a 256-bit perceptual signature for photo matching.

Show details
What it can support
Near-duplicate evidence relative to a particular reference and matching policy.
What it does not prove
Byte identity, safety category, copyright, or the truth of the reference.
Maturity
Public operationally oriented implementation.
Implementation
ThreatExchange PDQ C++ and integrations; use its quality and matching conventions.
Robustness to changes
Designed for perceptual copies; geometric edits and major content changes remain important tests.
Error caveats
Threshold/quality handling and corpus scale affect error rates. A reference label is not inherent in the hash.
License and access
ThreatExchange BSD license; repository notes file-level exceptions such as CImg.
Works offline
Yes; hashing does not require the ThreatExchange service.
Cost and dependencies
Local indexing; shared reference sets are a separate access dependency.
Privacy and rights
Fingerprint exchange can expose reference membership; sensitive hash sets need access controls.
Role for LegitPix
Practical optional near-duplicate matcher; preserve reference ID, score and threshold.
Uncertainty
No independent LegitPix benchmark or installed build verified.

Sources

Evidence type: software · Last reviewed 2026-10-06

Perceptual Matching· image

SSCD learned image-copy descriptors

Self-supervised image-copy model emits normalized embeddings for cosine/distance-based retrieval.

Show details
What it can support
Copy-like similarity to an indexed image, including some substantial visual alterations.
What it does not prove
Cryptographic equality, provenance chain, copyright, or semantic truth.
Maturity
CVPR 2022 research implementation with downloadable inference models.
Implementation
Meta SSCD TorchScript/PyTorch models; typically 512-dimensional descriptors, larger variant 1024.
Robustness to changes
Trained for image-copy transformations; test crops, overlays, recompression and screenshots on target data.
Error caveats
Repository's example precision/threshold is specific to DISC and its model; it is not a transferable probability.
License and access
Code MIT; verify downloaded weights and all dependencies separately.
Works offline
Yes after weights and dependencies are acquired; CPU or GPU.
Cost and dependencies
Model inference and vector index; target-domain normalization may be useful.
Privacy and rights
Embeddings retain linkability; external vector services transmit derived data.
Role for LegitPix
Higher-capacity candidate matcher, subject to local operating-point validation.
Uncertainty
Research environment uses older dependencies; no maintained SLA or commercial weight clearance inferred.

Sources

Evidence type: research · Last reviewed 2026-10-06

Provenance· image, audio, video, document

C2PA Content Credentials: signed manifests and hard bindings

Cryptographically sign claims/assertions and bind them to specified asset bytes or structured portions; actions and ingredients describe claimed history.

Show details
What it can support
Signature/claim integrity and covered-content consistency under an identified validation and trust policy.
What it does not prove
Scene truth, complete history, legal rights, or accuracy of every metadata field. Hard binding need not cover every byte.
Maturity
Published interoperable standard
Implementation
Open standard with multiple producer/validator implementations; manifests embedded or external. Specification 2.5 expands format and live-media coverage, adds linked-resource integrity, permits ML-DSA signatures, and clarifies binding and validation. These are published specification capabilities, not evidence that every implementation supports them.
Robustness to changes
Changed covered content fails old hard binding; a compliant editor may issue a new signed derivative referencing ingredients.
Error caveats
Missing, invalid, cryptographically valid and trusted are distinct states; signed assertions can be false.
License and access
Open royalty-free standard; specification content CC BY 4.0; software licenses separate.
Works offline
Conditional: local validation with asset/manifests/trust material; fresh revocation and remote ingredients may need network.
Cost and dependencies
Verification can be local/free; trusted production signing, timestamps and conformance may cost.
Privacy and rights
Manifests may expose creator, location or edit history; minimize before publication.
Role for LegitPix
Primary interoperable provenance layer alongside whole-file receipt hashing; retain excluded-range semantics.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed. Exact publication day and implementation-level 2.5 support were not established; check each feature against the chosen validator release.

Sources

Evidence type: standard · Last reviewed 2026-10-26

Provenance· image, audio, video

JPEG Trust framework

Framework for annotating and evaluating media provenance, integrity, attribution and trust under defined policies.

Show details
What it can support
Evidence can be evaluated against a stated trust profile and reported consistently.
What it does not prove
Universal truth, complete history or a legal ownership finding.
Maturity
Core standard published; additional parts still under development.
Implementation
ISO/IEC 21617 framework; independently verify implementations and profile support.
Robustness to changes
Media changes require renewed integrity/provenance evaluation; not itself a robust watermark.
Error caveats
Trust results depend on the selected profile and quality of evidence.
License and access
ISO standard copyright and implementation rights are separate; no reference-code license verified.
Works offline
Depends on validator, locally available assertions and external trust dependencies.
Cost and dependencies
Standards access and implementation costs require checking; none purchased.
Privacy and rights
Annotations can expose attribution or contextual data.
Role for LegitPix
Track standards alignment and explicit trust-policy reporting.
Uncertainty
Historical observations (not current contradictions): ISO lists edition 2 as published while the JPEG workplan still describes it as in development; use ISO for publication status. Full specification and implementation conformance not evaluated. On 19 October the ISO Part 4 ?browse=tc view showed dated stage 40.60, while the canonical URL and JPEG workplan returned older 40.20 snapshots. The newer dated ISO lifecycle record is used; cache disagreement is retained explicitly. On 26 October the JPEG workplan agrees with published Part 1 edition 2 and Part 4 DIS 40.60, resolving the earlier workplan disagreement; historical observations above are retained. Rechecked 2 November: the JPEG workplan still agrees on Part 1 edition 2 published, Part 2 FDIS 50.00, and Parts 3 and 4 DIS 40.60. Full specification and implementation conformance remain unevaluated.

Sources

Evidence type: standard · Last reviewed 2026-11-02

Provenance· image, audio, video, document

CAWG Identity Assertion

Separate signed identity assertion binds a credential holder and stated role to selected C2PA assertions, independent of the tool’s claim-signing identity.

Show details
What it can support
Control/verification of a specified digital identity and its signed relationship to selected assertions.
What it does not prove
Identity honesty, copyright ownership, complete authorship, or truth of depicted content.
Maturity
Ratified extension; ecosystem support varies
Implementation
X.509/COSE and identity-claims aggregation routes; compatible validators required.
Robustness to changes
Follows bound assertion integrity; no extra pixel robustness.
Error caveats
Display-name metadata and claim-generator certificate are not equivalent to verified creator identity.
License and access
W3C Patent Policy; specification reference code Apache 2.0 unless otherwise stated.
Works offline
Conditional on packaged credentials, trust roots and status; DID/status resolution can need network.
Cost and dependencies
Identity provider/CA integration may require accounts, verification and fees.
Privacy and rights
Opt-in identity can expose legal names, affiliations and account links permanently; minimize disclosures.
Role for LegitPix
Optional attributable evidence, not a requirement for ordinary users or anonymous documentation.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Provenance· image, audio, video

C2PA durable credentials: soft bindings and manifest recovery

Use watermark identifiers or content fingerprints to discover a stored manifest when embedded metadata is missing.

Show details
What it can support
Probabilistic association to candidate provenance records, followed by separate signature validation.
What it does not prove
Recovered association is not a successful hard binding to a transformed file or proof that all edits were logged.
Maturity
Standardized recovery architecture; deployment-dependent
Implementation
C2PA Soft Binding API and algorithm registry; fingerprint/watermark implementations and repositories vary.
Robustness to changes
Designed to reconnect transformed assets; success depends on each watermark/fingerprint and registered coverage.
Error caveats
False associations and absent records are possible; do not upgrade a soft match to byte-identical verification.
License and access
Open specification; algorithm/model/provider terms separate.
Works offline
Only with local detector/index and cached manifests; public repository discovery normally online.
Cost and dependencies
Registry storage, lookup, watermark licenses and uptime are separate dependencies.
Privacy and rights
Queries can reveal which content is being inspected; prefer hash/identifier-only or local lookup where supported.
Role for LegitPix
Optional candidate-discovery stage, with visibly weaker evidence state than exact-byte integrity.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Provenance· image, audio, video, document

CAI open-source validators and signers (c2pa-rs/c2patool/JS)

Implementation layer for reading, validating and authoring C2PA evidence; different language/runtime interfaces share core functionality.

Show details
What it can support
Results of the configured C2PA validator on supported formats, not a new kind of truth claim.
What it does not prove
The implementation is bug-free or conformant merely because it uses this SDK; no scene-truth verdict.
Maturity
Maintained beta SDK ecosystem; release status differs by component.
Implementation
Rust SDK, command-line c2patool and JavaScript/WASM ecosystem; public local validation possible. c2patool v0.28.2 updates its SDK dependency to 0.91.2 and fixes external-manifest handling during export and inspection. Rust SDK 0.91.3 release notes report SVG root namespace handling and Ed25519 certificate-parameter validation fixes; C FFI 0.91.3 adds a Builder label-setting API. These changes were not independently tested. Subsequent maintainer notes: 0.91.4 changes handling of SHA-1 timestamp imprints; 0.91.5 fixes rejection of valid MOV inputs; 0.91.6 updates the ECDSA stack and c2pa_raw_crypto to 0.2.0. c2patool 0.28.3 adopts SDK 0.91.6. No fixes were independently reproduced. Maintainer release 0.91.7 updates TLS dependency minimums and requires eMudhra distrust updates; c2patool 0.28.4 adopts SDK 0.91.7. This is dependency maintenance, not a new media-authenticity guarantee.
Robustness to changes
Format/feature support varies by release; unsupported evidence must be separate from invalid.
Error caveats
Pin release and configuration; preserve detailed validation codes rather than a single boolean.
License and access
Rust/c2patool MIT and Apache-2.0; JS project MIT; transitive licenses vary.
Works offline
Yes for local evidence; explicitly configure remote fetching, trust roots and status freshness.
Cost and dependencies
No SDK license fee; compute, storage, certificates and integration remain.
Privacy and rights
Local-only operation requires checking remote URL and telemetry behavior, not assuming it from “open source.”
Role for LegitPix
Preferred baseline for independent signature validation; no SDK or model was installed during this research.
Uncertainty
Release notes reviewed, not installed or independently reproduced. GitHub labels for the Rust and C FFI tags differ; a Latest tag for one component must not be generalized to every package. Pin the exact component/version and verify its supported specification features. The C FFI 0.91.7 release body repeats October 16 notes; its October 26 release date is taken from GitHub release metadata, not from that older body. The direct docs.rs 0.91.7 page could not be fetched.

Sources

Evidence type: software · Last reviewed 2026-11-02

Provenance· image, audio, video, document

C2PA trust lists, certificates and conformance

Validate certificate paths, timestamps and status against chosen trust roots; conformance evaluates a product implementation.

Show details
What it can support
A claim’s signer chains to the selected trust policy and applicable certificate conditions.
What it does not prove
A trusted signer is honest about every assertion, and conformance does not certify the depicted event.
Maturity
Operational governance infrastructure
Implementation
Official Conforming Products List, C2PA Trust List and TSA Trust List; legacy ITL remains distinguishable.
Robustness to changes
Independent of perceptual transformation robustness; validates cryptographic and governance evidence.
Error caveats
Unknown trust root is not the same as invalid signature. Legacy ITL acceptance depends on validator and validity period.
License and access
Public specifications/lists; onboarding and certificate contracts apply.
Works offline
Cached roots/status enable bounded offline checks; report freshness and unavailable status explicitly.
Cost and dependencies
Production signing may require evaluation, onboarding and certificate fees; reading standards is free.
Privacy and rights
Certificate and status lookups can expose activity; restrict unsolicited external fetching.
Role for LegitPix
Version trust policy and list snapshot in every verification receipt.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Reference Matching Service· audio

AcoustID reference-database lookup

Submit a Chromaprint fingerprint and duration to retrieve matching recording IDs and optional MusicBrainz metadata.

Show details
What it can support
A database association with a known recording.
What it does not prove
Complete audio provenance, correct rights ownership, or that absent matches are original.
Maturity
Operating public and commercial services.
Implementation
Documented v2 lookup API; application key required.
Robustness to changes
Inherits fingerprint and catalog-coverage limitations.
Error caveats
Crowdsourced metadata can be wrong or ambiguous; service score is not an authenticity probability.
License and access
Service software MIT; fingerprint/metadata database CC BY-SA 3.0; API terms separate.
Works offline
Hosted lookup no; self-hosted software/data is a separate engineering path.
Cost and dependencies
Public endpoint is noncommercial with 3 requests/second guideline; commercial use requires the commercial service.
Privacy and rights
Fingerprint/duration queries leave the device; submission is a separate data-sharing action.
Role for LegitPix
Optional external reference lookup with explicit user data-sharing authorization.
Uncertainty
No query, subscription, account creation or fingerprint submission performed.

Sources

Evidence type: software · Last reviewed 2026-10-06

Reference Matching Service· audio, video-audio-track, video

Commercial audio/reference recognition (ACRCloud representative)

Commercial API accepts audio or SDK-generated fingerprints and returns catalog/custom-reference recognition metadata.

Show details
What it can support
A provider-scoped reference match; some products return segment offsets.
What it does not prove
Authenticity, voice identity, ownership, or image/video-frame verification merely because the container is video.
Maturity
Documented commercial APIs and SDKs.
Implementation
ACRCloud identification endpoint with authenticated project credentials; custom buckets and product-specific metadata.
Robustness to changes
Product-specific matching robustness must be tested; avoid extrapolating provider marketing.
Error caveats
Catalog coverage, content duration and shared audio affect interpretation; provider scores need calibration.
License and access
Proprietary service/SDK terms; public client examples do not license the matching engine.
Works offline
Hosted identification requires network; docs mention offline project types, but licensing/deployment was not verified.
Cost and dependencies
Account, commercial plan, SDK and selected catalog/product.
Privacy and rights
Uploading audio or fingerprints transmits potentially private recording information.
Role for LegitPix
Optional adapter only after rights, disclosure, cost and data-processing review.
Uncertainty
No account, pricing quote, offline entitlement or accuracy validation obtained.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Reference Matching Service· image

Hosted private image-reference matching (TinEye MatchEngine representative)

Upload an authorized image collection, create provider fingerprints, and query that private index.

Show details
What it can support
Similarity to an item in the supplied reference collection.
What it does not prove
Public-web discovery, creator identity, ownership or scene truth.
Maturity
Commercial hosted matching service.
Implementation
MatchEngine API; catalog import and query workflow.
Robustness to changes
Provider advertises matching resized, cropped, color-changed and occluded variants; noisy mobile photos are a separate product focus.
Error caveats
Vendor robustness statements are not independent validation; threshold and reference-set composition matter.
License and access
Proprietary API/service terms.
Works offline
No public offline deployment established.
Cost and dependencies
Paid service, storage/index scope and vendor agreement.
Privacy and rights
Provider explicitly stores submitted reference images on its servers; approval and retention/access controls are required.
Role for LegitPix
Commercial alternative to self-hosted matching when data sharing is acceptable.
Uncertainty
No quote, deployment test or contractual privacy/security assessment performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Reference Matching Service· image

PhotoDNA known-image matching

Microsoft's perceptual signatures match copies against controlled reference sets of previously identified material.

Show details
What it can support
A candidate known-reference match within the approved program.
What it does not prove
General image authenticity, identity of a depicted person, or new/unknown content classification.
Maturity
Established restricted-access safety technology.
Implementation
PhotoDNA Cloud Service and approved tool-provider integrations.
Robustness to changes
Designed to find copies; precise current performance envelope not independently verified.
Error caveats
Reference coverage and policy determine meaning. No match cannot establish safety or authenticity.
License and access
Proprietary program/service terms.
Works offline
Not a publicly available general-purpose offline SDK; qualified access/integration must be established.
Cost and dependencies
Microsoft says free for qualified customers, law enforcement and tool providers; eligibility applies.
Privacy and rights
Highly sensitive safety workflows require specialized access, governance and lawful handling; do not build a private illegal-media reference set.
Role for LegitPix
Taxonomy/access-boundary reference only, not a default LegitPix integration.
Uncertainty
No eligibility, SDK access or service evaluation verified.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Reference Matching Service· audio, video

YouTube Content ID audiovisual reference matching

YouTube matches uploads to audio/visual reference files submitted by eligible rights holders.

Show details
What it can support
Platform-reported reuse of reference content.
What it does not prove
Authenticity, legal infringement, valid ownership, or comprehensive internet coverage.
Maturity
Deployed platform system with restricted rights-holder access.
Implementation
YouTube partner rights-management workflow, not a public authenticity API.
Robustness to changes
Audio and visual matching behavior is proprietary; do not assume all edited reuse is detected.
Error caveats
Erroneous claims occur and disputes exist. A Content ID claim is a rights-management event, not a legal adjudication.
License and access
Proprietary access terms; reference rights restrictions.
Works offline
No public offline implementation.
Cost and dependencies
Eligibility, exclusive rights, substantial catalog/need, geographic rights information and agreement.
Privacy and rights
Uploading media and enabling claims can have consequential publishing/rights effects.
Role for LegitPix
Documented external evidence if legitimately available; never use upload as an unapproved analysis workaround.
Uncertainty
No access or partnership assumed; no test upload made.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Registry· any-file, image, audio, video

Sigstore/Cosign/Rekor transparency evidence

Sign an artifact or receipt; bind short-lived certificates to an authenticated identity and log signing evidence in an append-only transparency service.

Show details
What it can support
Expected signer identity signed the committed artifact and evidence is included in the selected log.
What it does not prove
Signer assertions are true, a camera captured the file, copyright ownership, or a globally complete history.
Maturity
Deployed software-supply-chain infrastructure; media use is an adaptation
Implementation
Public Cosign blob signing/verification, Fulcio CA and Rekor; bundle-based evidence verification.
Robustness to changes
Exact artifact/receipt signature; edits require a new signed object.
Error caveats
A signed inclusion promise differs from an independently verified inclusion/consistency proof; verify expected identity and issuer.
License and access
Cosign Apache-2.0; service policies and other component licenses separate.
Works offline
Possible with complete bundle, trust root and required proof material; issuance/log submission online.
Cost and dependencies
Public-good services or self-hosted infrastructure; identity-provider availability and operational policies matter.
Privacy and rights
Public logs can permanently expose identity/email and digest. Never publish without explicit consent.
Role for LegitPix
Potential future auditable receipt infrastructure; not an automatic replacement for privacy-preserving local receipts.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: software · Last reviewed 2026-10-06

Reverse Search· image

Public reverse/visual image search (Google Lens representative)

Submit an image or image URL to discover visually related images and websites.

Show details
What it can support
A discovery lead linking visible content to indexed pages.
What it does not prove
Original source, earliest publication, truthful caption, license, or absence elsewhere.
Maturity
Operating consumer visual search.
Implementation
Google Lens browser/search interfaces; no general offline equivalent.
Robustness to changes
Can return similar subjects as well as pages with the same image.
Error caveats
Visual matches and AI summaries are leads, not verification. Index omissions, removed pages and changed context matter.
License and access
Proprietary service terms; results retain third-party rights.
Works offline
No.
Cost and dependencies
Hosted index, availability and interface limitations.
Privacy and rights
Image/URL is shared with Google; its help warns searched URLs may be stored.
Role for LegitPix
Opt-in source discovery followed by opening and corroborating the underlying page.
Uncertainty
No media upload or search benchmark performed; coverage is opaque.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Reverse Search· image

Reverse-copy web search (TinEye)

Fingerprint an image and search a crawled web index for copies and edited variants.

Show details
What it can support
An indexed occurrence or candidate modified copy.
What it does not prove
Creator, ownership, first publication, authenticity or originality after no match.
Maturity
Operating public reverse-image service and commercial API.
Implementation
TinEye web search; separate paid API product.
Robustness to changes
Finds cropped, edited, resized and partial copies; not usually different pictures of the same subject.
Error caveats
Indexed/crawl dates are not capture dates. Verify page context and compare the actual images.
License and access
Proprietary service/API terms; source-page content separately licensed.
Works offline
No for the public web index.
Cost and dependencies
Service availability, indexed coverage and API contract for automation.
Privacy and rights
Provider says search images are briefly retained and not added to its index; upload still requires authorization.
Role for LegitPix
Optional copy-discovery evidence with returned URL and observation time.
Uncertainty
No service query performed; retention claims are provider statements.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Semantic Retrieval· image, text, video-frames, video

Semantic image/text embeddings (CLIP-style retrieval)

Encode images and descriptions into a shared representation to retrieve semantically related content.

Show details
What it can support
Similarity in model representation; useful leads for scene or caption investigation.
What it does not prove
Same underlying asset, accurate caption, identity, event, place or capture time.
Maturity
Widely used retrieval building block; task validation still required.
Implementation
OpenAI CLIP or separately reviewed compatible implementations.
Robustness to changes
Semantic similarity may persist despite major visual differences, which is not evidence of copying.
Error caveats
Shared subject matter can produce high similarity for unrelated images. Model bias and out-of-domain behavior require evaluation.
License and access
Official CLIP code/model release uses MIT; third-party checkpoints may differ.
Works offline
Yes after weights are obtained.
Cost and dependencies
Model inference and vector search.
Privacy and rights
Do not use retrieval embeddings as an unreviewed biometric or sensitive-trait classifier.
Role for LegitPix
Discovery-only signal; never substitute its similarity for a copy hash.
Uncertainty
Original model card says deployment needs careful context-specific study.

Sources

Evidence type: research · Last reviewed 2026-10-06

Timestamp· any-file, image, audio, video

RFC 3161 trusted timestamps

A time-stamping authority signs a digest and time under a stated certificate/policy; RFC 5816 updates the protocol.

Show details
What it can support
The committed datum existed no later than the trusted timestamp, conditional on TSA/hash security.
What it does not prove
Original capture/creation time, authorship, truth, ownership, or public publication time.
Maturity
Long-established standard
Implementation
Standard TSA clients/servers; evidence can accompany a signature or receipt.
Robustness to changes
Exact digest commitment; changed bytes need a new timestamp or authenticated derivation.
Error caveats
Local computer time is not a trusted timestamp; check token policy, imprint, certificate and validity.
License and access
Open protocol; software licenses and TSA contracts separate.
Works offline
Verify stored token with trusted certificates/status evidence; issuance requires TSA or internal trusted service.
Cost and dependencies
TSA pricing varies; self-operated TSA creates its own trust burden.
Privacy and rights
Only digest is needed; digest disclosure can still link a known file to a timestamp request.
Role for LegitPix
Useful optional time anchor for local receipts; state “existed by” rather than “created at.”
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

  • RFC 3161 · August 2001; updated by RFC 5816

Evidence type: standard · Last reviewed 2026-10-06

Timestamp· any-file, image, audio, video

OpenTimestamps Bitcoin anchoring

Aggregate hash commitments into Bitcoin and retain a proof linking a file digest to a block.

Show details
What it can support
Committed bytes existed before a blockchain-anchored time bound under hash/chain assumptions.
What it does not prove
Who created the file, exact capture time, ownership, truthful content or that an incomplete proof is anchored.
Maturity
Public open-source protocol/tools
Implementation
Public browser/Python clients and calendar servers; .ots evidence file.
Robustness to changes
Exact hash binding; any changed file fails unless independently committed.
Error caveats
Pending calendar attestation is not completed Bitcoin proof; archive upgraded proof and verifier context.
License and access
Python client LGPLv3-or-later; other implementations have separate terms.
Works offline
Complete proofs can verify locally with required Bitcoin chain data; Python client documents local Bitcoin Core.
Cost and dependencies
Public calendars advertise free use without registration; direct anchoring has network fees. Confirmation may take hours.
Privacy and rights
Local hashing avoids raw upload; commitments and network metadata still have correlation risk.
Role for LegitPix
Optional independent existence anchor, not required for basic local byte receipts.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: software · Last reviewed 2026-10-06

Video Matching· video

TMK+PDQF temporal video matching

Temporal Match Kernel combines visual frame features into a fixed-size video signature.

Show details
What it can support
Whole-video similarity to a known reference under the implementation's alignment/scoring model.
What it does not prove
Framewise integrity, exact clip inclusion, or authenticity of the depicted event.
Maturity
Public Meta video-matching implementation.
Implementation
ThreatExchange TMK+PDQF; decoding dependencies and optional vector indexing.
Robustness to changes
Optimized for videos of comparable/same length; frame-feature limitations also apply.
Error caveats
Official comparison says TMK is not the subsequence/clip-matching alternative. Sampling and temporal edits can change results.
License and access
ThreatExchange BSD with documented file/dependency exceptions.
Works offline
Yes.
Cost and dependencies
Video decoding, signature storage and search.
Privacy and rights
Store authorized reference signatures with source provenance.
Role for LegitPix
Optional whole-video relationship evidence, distinctly labeled from vPDQ clip matches.
Uncertainty
No benchmark/runtime verification performed for LegitPix.

Sources

Evidence type: software · Last reviewed 2026-10-06

Video Matching· video, image

vPDQ frame-overlap and clip matching

Hash sampled frames using PDQ and compare shared similar frames, permitting subsequence/segment matches.

Show details
What it can support
A candidate shared visual segment or frame relationship.
What it does not prove
Identical ordering/content of all frames, audio integrity, or event authenticity.
Maturity
Public Meta implementation and Python integrations.
Implementation
ThreatExchange vPDQ C++/Python, video libraries; FAISS integration documented.
Robustness to changes
Supports clip/subsequence reuse; inherits PDQ's image-level strengths and weaknesses.
Error caveats
Shared title cards, static frames or repeated scenes are weak evidence alone; report coverage and time spans.
License and access
ThreatExchange BSD with dependency/file-level exceptions.
Works offline
Yes after build; official README says wheels are not distributed.
Cost and dependencies
Variable-length fingerprints; decoding, indexing and candidate verification costs.
Privacy and rights
Frame-level reference matching remains linkable derived data.
Role for LegitPix
Segment relationship evidence plus frame matches, not a binary authenticity verdict.
Uncertainty
Coverage, sampling rate and thresholds must be recorded and validated.

Sources

Evidence type: software · Last reviewed 2026-10-06

Watermark· audio, video

ATSC A/334 audio watermark emission

Standardized VP1 watermark embeds packets in PCM audio for broadcast and redistribution workflows.

Show details
What it can support
Recovery of an embedded broadcast payload by a compatible extractor.
What it does not prove
AI origin, scene truth, trustworthy authorship or an unmodified recording.
Maturity
Published standard
Implementation
ATSC 3.0 audio watermark format; specialist embedder/extractor required.
Robustness to changes
Designed to carry data through audio distribution, compression and channel mixing; validate the intended path.
Error caveats
Emission is optional. Unmarked and unsupported audio are not evidence of synthesis.
License and access
Publicly downloadable specification; potential patent licenses explicitly noted. Not an open-source software grant.
Works offline
Signal extraction can be local in an implementation; content recovery may require a service.
Cost and dependencies
Implementation and service pricing not verified.
Privacy and rights
Associated identifiers or viewing measurements require appropriate data controls.
Role for LegitPix
Specialist broadcast recognition option, outside the initial photo-receipt core.
Uncertainty
No implementation, patent terms or detector performance independently tested.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Watermark· video

ATSC A/335 video watermark emission

Broadcast video watermark carries ancillary data through image pixels for receiver content recovery.

Show details
What it can support
Extraction of an associated broadcast payload from compatible marked video.
What it does not prove
Indelible marking, tamper resistance, authentic footage or an ownership determination.
Maturity
Published standard
Implementation
ATSC 3.0 format, used with A/336 content-recovery signaling.
Robustness to changes
Designed for compression-rate changes, codec transcoding and HDMI redistribution; explicitly not tamper-resistant.
Error caveats
Optional emission and unsupported signal conditions make absence inconclusive.
License and access
Downloadable specification with patent-rights warning; implementation rights need separate review.
Works offline
An extractor can process video locally; resolving payloads may need broadband services.
Cost and dependencies
Implementation, patent and service costs unverified.
Privacy and rights
Content IDs can support viewing linkage; review receiver telemetry separately.
Role for LegitPix
Document the distinction between content-recognition signaling and authenticity evidence.
Uncertainty
No implementation or detection benchmark performed.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Watermark· audio

Audio echo hiding

Encode bits in low-level echo characteristics. Perceptual masking helps hide added signal, with recovery based on the echo structure.

Show details
What it can support
Recovery of an expected audio payload under the configured channel and decoder.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established algorithm family
Implementation
Classical published signal-processing family; no maintained original production package verified.
Robustness to changes
Channel reverberation, filtering and duration influence recoverability; no universal robustness number established.
Error caveats
Presence thresholds and synchronization matter. A hidden message is not authenticated merely because it decodes.
License and access
Publication does not establish an implementation or patent license; no legal status assessment.
Works offline
In principle yes, with conventional CPU signal processing.
Cost and dependencies
Implementation/integration work; no neural-weight service required.
Privacy and rights
Local implementation possible; payload identifiers should minimize personal data.
Role for LegitPix
Historical/classical audio baseline; compare against modern local tools before considering deployment.
Uncertainty
Primary published mechanism verified; current production readiness, perceptual tests and error rates not independently reproduced.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· audio

Audio phase coding

Encode data through controlled phase changes while preserving relative phases to limit audible distortion. Receiver synchronization and parameter agreement are essential.

Show details
What it can support
Recovery of an expected audio payload under the configured channel and decoder.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established algorithm family
Implementation
Classical published signal-processing family; no maintained original production package verified.
Robustness to changes
Original experiments concern a specified signal/decoder setup; modern codecs, time changes and recording channels require new testing.
Error caveats
Presence thresholds and synchronization matter. A hidden message is not authenticated merely because it decodes.
License and access
Publication does not establish an implementation or patent license; no legal status assessment.
Works offline
In principle yes, with conventional CPU signal processing.
Cost and dependencies
Implementation/integration work; no neural-weight service required.
Privacy and rights
Local implementation possible; payload identifiers should minimize personal data.
Role for LegitPix
Historical/classical audio baseline; compare against modern local tools before considering deployment.
Uncertainty
Primary published mechanism verified; current production readiness, perceptual tests and error rates not independently reproduced.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· audio

AudioSeal

Neural waveform watermark with sample-level localization and optional 16-bit message; detection and message extraction are separate outputs.

Show details
What it can support
Detected portions carry an AudioSeal-style mark; optional payload identifies configured source/version.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Open implementation with streaming
Implementation
Public generator/detector and training code; v0.2 adds streaming.
Robustness to changes
Tests cover compression, re-encoding and noise; speech/rate/channel assumptions matter.
Error caveats
Short/mixed clips and threshold choice matter; 16-bit identifiers are not globally unique or cryptographic authentication.
License and access
MIT code and weights since April 2, 2024.
Works offline
Yes after weights; CPU/GPU performance depends on length.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Candidate local audio localization adapter; preserve per-segment scores and expected-message check.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· audio, video-audio-track, video

audiowmark (spectral patchwork)

Traditional spectral patchwork audio marking with error correction and optional keyed operation.

Show details
What it can support
Recovery of an expected 128-bit identifier without original audio.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Maintained open-source tool
Implementation
C++ CLI; companion video wrapper marks the audio track, including HLS use.
Robustness to changes
Project reports usual recovery after MP3/Ogg at 128 kbps or above in its tests; other channels require evaluation.
Error caveats
Short duration and recording channel affect recovery; project recommends hash/HMAC-style payloads.
License and access
GPLv3-or-later; assess distribution/linking obligations.
Works offline
Yes, conventional CPU/FFT dependencies; no neural weights.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Non-neural local audio baseline with explicit copyleft review.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: software · Last reviewed 2026-10-06

Watermark· image

ChunkySeal

Scaled-up learned watermark emphasizing larger payload capacity.

Show details
What it can support
Recovery of up to 1024 embedded bits under tested conditions.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Model card/checkpoint in Meta VideoSeal repository.
Robustness to changes
Paper reports fourfold capacity over 256-bit baseline with measured quality/robustness; not a bound for arbitrary edits.
Error caveats
Capacity is not authentication strength. Reserve payload for opaque identifier or authenticated data.
License and access
MIT code and models per repository.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
Larger local model; benchmark memory and throughput.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Capacity-oriented research candidate; signed metadata can usually carry richer context without pixel payload pressure.
Uncertainty
Paper demonstrates image capacity. A video-capable repository/card is not independent video-performance validation; no model run here.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, audio

Digimarc Validate Media and watermark-linked recovery

Commercial covert watermark identifier linked to content ownership/provenance records and detection services.

Show details
What it can support
Detected content identifier associates media with a provider-managed record.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Commercial vendor ecosystem
Implementation
Commercial image/audio protection and validation API; public MIT-licensed beta Chrome extension recovers supported watermark-linked C2PA via SBR endpoints.
Robustness to changes
Vendor describes robust identifiers; detailed independent operating curves and coverage not verified here.
Error caveats
A vendor record is an assertion, not automatic copyright ownership or factual authenticity.
License and access
Validate services are proprietary; separate public browser-extension code is MIT.
Works offline
Browser C2PA checks exist; watermark-linked recovery uses external Digimarc/Adobe SBR services. General offline commercial detector access not verified.
Cost and dependencies
Contact vendor; public unit price not verified.
Privacy and rights
Browser extension processes media URLs/bytes and uses external SBR endpoints when enabled; broad site permissions and automatic scanning require careful consent.
Role for LegitPix
Vendor adapter only when a real user need and access justify cost; not baseline dependency.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Watermark· image

Fragile spatial/LSB authentication watermarks

Embed authentication bits in pixels, often blockwise low-bit data. Fragility intentionally reveals changes; hierarchical dependencies can address weaknesses of independent blocks.

Show details
What it can support
Consistency with an authenticated marked version, sometimes localized to blocks.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established family; legacy research examples
Implementation
Published public-key and hierarchical designs; no maintained production package verified.
Robustness to changes
Compression, resampling and ordinary edits may intentionally invalidate the mark; unsuitable as a durable internet identifier.
Error caveats
Change detected does not identify malicious intent. Historic independent-block schemes have published forgery weaknesses.
License and access
Paper rights retained; implementation and patent position not established.
Works offline
Yes; basic image processing plus cryptographic key management.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Controlled lossless pipelines; distinguish tamper-localization evidence from robust attribution.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, video, audio

Google SynthID for images, video and audio

Provider-integrated imperceptible media marks, using a learned image encoder/decoder and modality-specific systems. Visual marking can be post-hoc inside a generation workflow.

Show details
What it can support
Supported content carries a recognized SynthID mark from Google or supported partners; generated or AI-edited content may be marked.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Deployed provider system; public SynthID Detector announced globally in English on 2026-10-07.
Implementation
Public SynthID Detector for supported images, video and audio. Google names OpenAI, NVIDIA and Kakao as supported partners; Apple support is forthcoming. Built-in Google product checks have their own documented coverage.
Robustness to changes
Google documents common-edit resilience; technical paper evaluates external SynthID-O, not every production detector/model.
Error caveats
No detected mark does not mean human-made; marks may indicate editing rather than wholly synthetic creation.
License and access
Proprietary media models; public text code does not license image/audio/video detectors.
Works offline
No public standalone media detector/weights verified.
Cost and dependencies
Service/account and regional availability dependencies; no local licensing or API quote established.
Privacy and rights
Service checks require media upload; authorization and retention review needed. No media uploaded in this research.
Role for LegitPix
Optional provider-specific external check with explicit provenance scope and privacy gate.
Uncertainty
Public availability is a vendor announcement, not an independent accuracy or account-access test. No media uploaded; no public standalone media weights or general third-party detector API verified. Do not treat forthcoming Apple support as released.

Sources

Evidence type: vendor/research · Last reviewed 2026-10-12

Watermark· image

HiDDeN

Joint neural encoder/decoder learns imperceptible payload embedding with simulated distortions and adversarial image-quality training.

Show details
What it can support
Recovery of a watermark message with the matching trained decoder.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official jirenz/HiDDeN Torch/Lua code; README says WIP and pretrained models coming soon.
Robustness to changes
Paper tests blur, dropout, cropping and JPEG approximations; augmentation-trained robustness is distribution-specific.
Error caveats
Calibrate on unmarked negatives and transformed positives; decoding bits alone is not proof of a valid expected payload.
License and access
MIT code; dataset rights separate.
Works offline
Possible; legacy dependencies and training/weights acquisition make this a baseline rather than turnkey integration.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Historical learned-watermark baseline, not preferred production adapter.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, video

IMATAG watermarking and authenticity

Commercial image/video marking; default key-based detection and optional payload modes support tracing and association.

Show details
What it can support
Match to a registered watermark/key and associated customer record.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Commercial deployment
Implementation
Documented REST API v4; SaaS/API/on-premise offerings.
Robustness to changes
Robustness is configuration/service-specific; no independently verified general rate quoted.
Error caveats
Default detection needs the proper key/registration; absent match is not evidence of non-AI origin.
License and access
Proprietary commercial software/service.
Works offline
On-premise offered; exact detector, telemetry and license conditions unverified.
Cost and dependencies
Authenticity/Leaks pricing on request; Monitor advertised from EUR 299/month, not an authenticity API price.
Privacy and rights
Vendor workflow may require registration and uploaded media; no account created or content submitted.
Role for LegitPix
Optional customer-specific traceability connector, subject to licensing/privacy review.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor/software · Last reviewed 2026-10-06

Watermark· image

invisible-watermark (DWT/DCT, SVD, RivaGAN port)

Practical blind image watermark library covering classical transform variants and a learned model.

Show details
What it can support
Recovery of a caller-selected payload, subject to verification.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
ShieldMnt Python package and CLI; experimental CPU-only implementation.
Robustness to changes
Repository reports JPEG/noise/brightness successes, weaker geometric resilience and poor homogeneous screenshots/posters.
Error caveats
Authors warn even untouched media may not decode perfectly. README toy tests are not deployment-wide accuracy estimates.
License and access
MIT repository.
Works offline
Yes; basic methods CPU-feasible; learned/SVD routes slower.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Low-cost baseline only; not a generic watermark scanner or production reliability claim.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: software · Last reviewed 2026-10-06

Watermark· image, video

PixelSeal

Adversarial-only training and high-resolution adaptation improve the quality/robustness trade-off of image/video watermarks.

Show details
What it can support
Expected 256-bit payload recovery with the released model.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Checkpoint and inference in Meta VideoSeal repository.
Robustness to changes
Authors evaluate many visual transformations; JND attenuation and temporal pooling address high resolution/video.
Error caveats
“State of the art” is benchmark-relative; public detector/encoder availability is not unforgeability.
License and access
MIT code and models per repository.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Promising permissive local baseline; benchmark on LegitPix media before adoption.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, audio, video

Quantization index modulation (QIM/DC-QIM)

Encode bits through quantizer choices in samples or transform coefficients. Distortion compensation adjusts the robustness/capacity trade-off.

Show details
What it can support
A payload matching an expected codebook/key was recovered.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established algorithm family
Implementation
Algorithm family implementable in image/audio/video pipelines.
Robustness to changes
Provable results concern stated bounded-distortion or probabilistic channel models; real codecs and synchronization errors require measurement.
Error caveats
Theoretical optimality does not imply robustness to arbitrary editing or forged attribution.
License and access
Primary paper does not supply a universal software/patent grant.
Works offline
Yes; no neural weights required.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Optional method-specific evidence; record detector/version, key scope, threshold, payload check and limitations.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Reversible watermarking/data hiding

Embed payload while retaining enough information to reconstruct the original pixels exactly under the method’s conditions; difference expansion is one example.

Show details
What it can support
Exact recovery of the pre-embedding image when reversibility conditions hold.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established research family
Implementation
Tian difference expansion; histogram/prediction variants form a larger family.
Robustness to changes
Reversible does not mean durable: lossy transformations can destroy payload and exact recovery.
Error caveats
Recoverability alone authenticates neither the payload nor the creator; add cryptographic checks.
License and access
Paper rights retained; no production software license verified.
Works offline
Yes; CPU-friendly algorithms exist.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Archival/controlled workflows where altering original pixels is unacceptable after extraction.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· video

RivaGAN

Attention-based video encoder/decoder with adversarial networks balancing quality and robustness.

Show details
What it can support
A known 32-bit watermark message recovered from supported video.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official DAI-Lab Python implementation exposes train, encode and decode.
Robustness to changes
Authors report robustness to common processing; no universal current-codec guarantee.
Error caveats
Calibrate on unmarked negatives and transformed positives; decoding bits alone is not proof of a valid expected payload.
License and access
MIT code; training datasets separate.
Works offline
Possible with trained model; old Python requirements increase maintenance cost.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Legacy video-watermark baseline; image ports must be evaluated separately.
Uncertainty
Official research targets video; image-port coverage is separately catalogued under invisible-watermark. No runtime reproduction.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Semi-fragile authentication and SARI

Accept a defined class of benign processing while rejecting other changes; SARI combines localization with approximate recovery data.

Show details
What it can support
Media conforms to the scheme’s allowed-change policy.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research family
Implementation
Lin–Chang JPEG semi-fragile and SARI prototypes; current supported distribution unverified.
Robustness to changes
Representative design tolerates JPEG quantization only to a predetermined quality boundary; other transformations need separate evaluation.
Error caveats
Benign/malicious is an engineered policy, not a semantic judgment; approximate recovery is not recovered original evidence.
License and access
No current software license verified.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Specialized controlled workflows; avoid a generic “not manipulated” verdict.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· video

Session-specific forensic video watermarking (NexGuard)

Embed a deliberately assigned identifier into each distributed session/copy. This “forensic fingerprint” is not a content-derived perceptual fingerprint; tracing requires the issuance record.

Show details
What it can support
Recovered identifier corresponds to an issued copy or session.
What it does not prove
Who personally leaked the copy, scene truth, ownership, or that a subscriber intentionally acted.
Maturity
Commercial distribution infrastructure
Implementation
NAGRA NexGuard Streaming integrates server-side with CDNs for live and VOD.
Robustness to changes
Vendor robustness claims require service-specific validation; collusion and capture channels belong in the threat model.
Error caveats
Do not turn a session match into an accusation against a person without corroboration.
License and access
Proprietary commercial service.
Works offline
Not a public standalone detector; depends on vendor/session records.
Cost and dependencies
Negotiated enterprise service and CDN/distribution integration.
Privacy and rights
Session identifiers can identify viewers; treat mappings as restricted personal data.
Role for LegitPix
Separate leak-tracing use case, outside baseline authenticity receipt scope.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor · Last reviewed 2026-10-06

Watermark· audio

SilentCipher

Neural audio watermark with psychoacoustic thresholding and compression-aware training; 40-bit messages.

Show details
What it can support
Recovery of a recognized audio payload.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Sony Python inference for 16 kHz and 44.1 kHz; downloadable models; CPU/CUDA paths.
Robustness to changes
Compression and audio distortions evaluated; crop-aware decoding has additional runtime cost.
Error caveats
Changing embedding strength can alter detection accuracy; “inaudible” depends on content/listening conditions.
License and access
MIT code; model distribution terms should be pinned separately.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
High-sample-rate audio comparison candidate, subject to model-license review.
Uncertainty
Repository still describes training-code release as forthcoming; no independent runtime reproduction.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Stable Signature

Fine-tune a latent-diffusion decoder so generated output carries a fixed learned watermark.

Show details
What it can support
Consistency with a configured watermarked decoder/key.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Meta training, extraction and evaluation code; 48-bit extractor checkpoints.
Robustness to changes
Alternative extractors trade image quality against blur/rotation robustness; transformed outputs need evaluation.
Error caveats
Watermarked-model attribution is conditional on secret/control assumptions; released keys/models are not exclusive author credentials.
License and access
Predominantly CC-BY-NC; some ldm/taming code MIT; base-generator licenses separate.
Works offline
Yes for licensed model owners; training/generation typically GPU-heavy.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Generator-side research comparator, not a scanner for all diffusion images.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, video

StegaStamp

Neural image watermark designed for print/photograph recovery; encodes small messages plus error correction.

Show details
What it can support
Recovery of an embedded message through the evaluated physical imaging channel.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official TensorFlow code, weights and image/video detector; default 100 coded bits yield 56 payload bits.
Robustness to changes
Physical print-camera and perspective/photometric changes evaluated; detector size sensitivity and noisy region detection documented.
Error caveats
Calibrate on unmarked negatives and transformed positives; decoding bits alone is not proof of a valid expected payload.
License and access
MIT repository; check bundled dependencies.
Works offline
Yes after weights; legacy TensorFlow 1.13 environment may need isolation.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Optional durable locator for print/recapture scenarios, never evidence that the photographed scene itself occurred.
Uncertainty
Detector training code is not fully included; runtime and real-world recapture performance not reproduced.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Transform-domain spread-spectrum watermarking

Distribute keyed low-amplitude signal across perceptually important spectral components; correlation tests retrieve presence or payload.

Show details
What it can support
A signal consistent with the expected watermark/key is present.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Established algorithm family
Implementation
Cox et al. foundational DCT/spectral design; many derivatives.
Robustness to changes
Paper evaluates signal processing and geometric changes; geometric robustness assumes successful registration and available original in its design.
Error caveats
Correlation threshold, key count and original access affect false matches; do not transfer one design’s guarantees to the whole family.
License and access
Method publication is not a blanket implementation/patent license.
Works offline
Yes; CPU implementations feasible; original/reference requirements vary.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Optional method-specific evidence; record detector/version, key scope, threshold, payload check and limitations.
Uncertainty
Source demonstrates image watermarking and proposes generalization to audio/video; those modalities need specific implementations and evidence.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

TrustMark

Learned arbitrary-resolution image watermark with configurable error correction and public encoder/decoder.

Show details
What it can support
Expected payload recovery; can locate a separate C2PA manifest.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Open implementation with C2PA integration
Implementation
Adobe implementation includes Python, JavaScript and Rust routes; models downloaded on first use.
Robustness to changes
Targets common visual transformations. Project also publishes a removal model, so robustness must not be read as irreversible or unremovable.
Error caveats
A very recent single preprint reports ECC-only presence false alarms of 5.9–15.3% for evaluated variants. Verify expected payload, not merely successful ECC decode.
License and access
MIT explicitly covers code and model files.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Strong local-integration candidate after calibrated expected-payload validation; separate soft association from hard integrity.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, video

VideoSeal

Post-hoc image/video watermark using efficient temporal propagation and aggregation.

Show details
What it can support
Expected payload recovered from visual frames.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Open model plus evaluation tooling
Implementation
Meta PyTorch library; v1.0 256-bit stable model; streaming inference and evaluation scripts.
Robustness to changes
Evaluations cover visual alterations and video compression; visual model does not authenticate audio track.
Error caveats
Calibrate on unmarked negatives and transformed positives; decoding bits alone is not proof of a valid expected payload.
License and access
MIT model/code; third-party baselines have their own terms.
Works offline
Yes; PyTorch/codec dependencies; streaming avoids loading long videos fully.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Candidate local video adapter. Keep visual and audio evidence separate.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, video, audio

Visible overlays and audible disclosure marks

Human-perceptible logo, text, pattern or audible cue accompanies content. Can carry an identifier pointing to a separate receipt.

Show details
What it can support
A displayed attribution or disclosure claim exists.
What it does not prove
Who applied the label, whether its claim is correct, file integrity, scene truth or legal ownership.
Maturity
Widely deployed family
Implementation
Ordinary image/video compositors and audio editors; no specialized detector.
Robustness to changes
Visibility can survive sharing but may be obscured or lost in derivatives; no cryptographic resistance.
Error caveats
A forged or copied label looks legitimate; absence has no AI-origin meaning.
License and access
Implementation-specific; no universal license.
Works offline
Yes; low-cost conventional processing.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Disclosure/wayfinding only. Do not elevate an overlay to verified provenance.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: standard · Last reviewed 2026-10-06

Watermark· image

Watermark Anything (WAM)

Localized neural watermarking can embed different messages in image regions and recover location/payload maps.

Show details
What it can support
A region carries a detected message; may support tracing composites.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Archived research repository
Implementation
Meta code and SA-1B checkpoint; original paper COCO checkpoint separate.
Robustness to changes
Designed for localized crops/composites and transformed regions; mask accuracy depends on scale and calibration.
Error caveats
Region localization is probabilistic; detector output is not an exhaustive edit map.
License and access
MIT code/new SA-1B weights; original COCO weights CC-BY-NC 4.0.
Works offline
Yes with selected checkpoint; maintenance responsibility is local.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Research adapter for localized provenance; choose license-compatible checkpoint explicitly.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· audio

WavMark

Neural 32-bit watermark per one-second audio window; public wrapper reserves 16 bits for detection and 16 for user payload.

Show details
What it can support
Recovery of marked payload in the evaluated audio pipeline.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Public PyTorch implementation; documented mono 16 kHz input.
Robustness to changes
Authors test ten alterations and report over 38 dB SNR/4.3 PESQ; perceptual metrics are not universal inaudibility proof.
Error caveats
The 1/2^16 pattern argument is per candidate under random-bit assumptions, not a calibrated file-level false-positive rate across many windows.
License and access
MIT repository; separately confirm redistributed model asset terms.
Works offline
Yes, after obtaining dependencies and weights; local performance untested.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Audio baseline; report search-window multiplicity and calibrate real negatives.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Gaussian Shading

Distribution-preserving latent sampling encodes a message; diffusion inversion retrieves the latent and message.

Show details
What it can support
Expected latent watermark payload match under configured model/key.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official PyTorch implementation includes encrypted randomization option.
Robustness to changes
“Performance-lossless” concerns the ideal latent distribution, not resistance to every edit or zero pixel distortion.
Error caveats
Distribution and cryptographic assumptions matter; implementation options can weaken theoretical guarantees.
License and access
MIT code; base model licenses separate.
Works offline
Yes with model weights/key; GPU recommended.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Research comparator for provenance at generation, not universal AI detection.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Pseudorandom-code diffusion watermark (PRC)

Use pseudorandom error-correcting codes to select diffusion latents, combining cryptographic indistinguishability assumptions with watermark recovery.

Show details
What it can support
A matching keyed watermark is detected under stated model/security assumptions.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official PRC-Watermark code; public example exposes binary detection; long-message interface described as forthcoming.
Robustness to changes
Paper reports robustness on Stable Diffusion 2.1 and 512-bit robust payload experiments; repository interface is narrower.
Error caveats
Undetectability means an outsider cannot distinguish watermarking under assumptions; it does not mean a public detector can identify every AI image.
License and access
MIT code; diffusion weights/licenses separate.
Works offline
Yes with weights and key; diffusion inversion/generation compute required.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Security-oriented research comparator; distinguish theoretical properties, tested attacks and released features.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image

Tree-Ring watermark

Choose diffusion initial noise containing a structured Fourier-domain key; invert an image with a compatible model to test for the key.

Show details
What it can support
Image is statistically consistent with the configured latent watermark.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Research implementation
Implementation
Official PyTorch/diffusers research implementation; version-sensitive DDIM inversion.
Robustness to changes
Paper evaluates geometric and photometric distortions; robustness depends on inversion and model matching.
Error caveats
A statistical latent match is not cryptographic provenance or human identity.
License and access
MIT code; generator/model/data licenses separate.
Works offline
Yes with compatible diffusion weights and key; GPU/inversion compute significant.
Cost and dependencies
No quoted license fee; local compute and integration costs.
Privacy and rights
Prefer local processing; use opaque identifiers rather than personal data. Embedding does not establish rights.
Role for LegitPix
Generator-integrated research family; do not describe as ordinary metadata fingerprinting.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: research/software · Last reviewed 2026-10-06

Watermark· image, audio

OpenAI Content Provenance verifier (C2PA + SynthID)

Provider verifier reports separate provenance signals: image C2PA and SynthID; supported audio SynthID.

Show details
What it can support
Recognized OpenAI-origin watermark or OpenAI-issued AI-generation credentials, as separately reported.
What it does not prove
Scene truth, legal ownership, complete history, or absence of AI. A recovered mark is not itself a digital signature.
Maturity
Public verification service/API
Implementation
Public web verifier and POST /v1/content_provenance_checks API; no top-level merged outcome in documented response.
Robustness to changes
Metadata may be lost while a watermark survives; degraded marks and legacy outputs may be missed.
Error caveats
Not a general-purpose AI detector. C2PA not_detected can coexist with a valid third-party manifest; read validation_state and issuer independently.
License and access
Proprietary service under applicable OpenAI terms; no local media model license.
Works offline
No public offline verifier for SynthID confirmed; generic C2PA can be independently checked.
Cost and dependencies
API account/service dependencies; per-check pricing not verified.
Privacy and rights
Uploads transmit media to OpenAI; require task-specific authorization. No upload made.
Role for LegitPix
Separate provider-origin evidence adapter; preserve raw per-signal states, never reduce to real/fake.
Uncertainty
Research-only review; no media uploaded, model run or independent reproduction performed.

Sources

Evidence type: vendor/software · Last reviewed 2026-10-06

Evaluate before integrating

An authored test plan for LegitPix. These checks have not been run against the cataloged tools. Use deployment-specific measurements rather than transferring a paper's headline result.

State the decision first

Define whether the task is exact identity, near-copy retrieval, mark presence, payload attribution, signer validation, partial-edit localization, face manipulation, full synthesis or contextual verification. These require different ground truth and cannot share one accuracy label.

Build a rights-cleared test collection

Use media with documented permission, source and transformations. Separate real captures, ordinary edits, fully generated media, AI-assisted edits, composites and re-recordings. Include difficult legitimate cases. Keep subjects, capture devices, source videos and near-duplicates from leaking across train, validation and test splits.

Test the actual distribution path

Measure results on originals and ordinary deployment transformations: resizing, compression, format conversion, cropping, screenshots, metadata loss, video transcoding, frame-rate changes, excerpting, audio compression and recording conditions. Evaluate individual and combined transformations without presenting the results as removal instructions.

Report errors at the chosen operating point

Report true-positive rate at a predeclared false-positive rate, precision at plausible prevalence, false-negative rate, uncertainty intervals and abstention coverage. Include per-domain and low-quality-input results. AUROC, balanced-set accuracy and vendor headline numbers do not establish a suitable production threshold.

Evaluate the right unit

For matching, report candidate precision/recall at index scale and verify the reference link. For marks, separate present/absent detection, payload recovery, wrong-key or wrong-identity attribution, and perceptual quality. For provenance, test binding, signatures, trust decisions, time, revocation and missing ingredients independently. For localization, measure regions and time spans as well as whole-file classification.

Treat dependencies as evidence

Pin code, model weights, preprocessing, reference-index version and license. Measure CPU/GPU time, memory, file limits, failures and network calls on the intended platform. Test offline behavior explicitly. GPU inference demonstrated in a paper is not evidence that a model runs acceptably in a browser.

Validate updates before replacement

Maintain held-out, time-separated cases for new generators and devices. Track drift and regression, including false accusations on legitimate work. Correlated detectors are not independent witnesses; do not multiply their scores or average them into a probability without a validated fusion model.

Illustrative calculation, not a measured detector result: in 10,000 files with 1% target prevalence, 90% sensitivity and 1% false-positive rate yield about 90 true alerts and 99 false alerts. Only about 48% of alerts would be correct. The real prevalence, threshold and error rates must be measured for the deployment.

Measurement references

  • NIST AI 100-4 November 2024

    Broad taxonomy and measurement guidance. Provenance and detection provide complementary evidence; low false-positive operating points and application context matter. The protocol in this library is an authored recommendation, not a NIST certification.

  • NIST Open Media Forensics Challenge Historical evaluation page, accessed 2026-10-06

    Separates manipulation detection, deepfake detection and localization tasks. Documents ROC-based and localization metrics. The page says the evaluation has finished; archived benchmark results are not current deployment guarantees.

  • Deepfake-Eval-2024 v5, 27 May 2026

    Image, audio and video benchmark of media circulating in 2024. Authors report substantial performance drops relative to earlier benchmarks. Its collection and labels do not establish every product's current accuracy or future generalization.

  • NIST GenAI evaluation program Updated 7 May 2026; accessed 2026-10-06

    Current official program entry point links image, text and code challenge work. Use published task plans and current eligibility terms; no participation or dataset access was requested for this library.

  • Watermarks and Fingerprints as Soft Bindings for Content Provenance v1, 2 October 2026; very recent preprint

    Original Pictures Technologies authors compare openly available watermark and fingerprint configurations under one protocol. Useful for evaluation design and payload-versus-presence distinctions; not peer-reviewed evidence of a universal winner. Findings need independent replication on the intended workload.

Scope, gaps and maintenance

Catalog schema, record IDs, CSV/JSON agreement, embedded data and search/filter/export logic were checked locally. Full browser rendering and actual browser downloads could not be verified in this environment; the HTML remains a self-contained reference with all records readable without JavaScript.

Baseline research completed 6 October 2026; bounded primary-source maintenance reviews completed 12, 19 and 26 October and 2 November 2026. Source-specific dates show which references were revisited. Implementations were not installed or benchmarked, commercial terms were not negotiated, and no personal media was uploaded. Recheck the exact version and licenses before adoption.

  • Proprietary internals, closed detector keys, private reference databases, unpublished methods and contractual service terms cannot be fully inspected.
  • Representative implementations cover the taxonomy; this is not an exhaustive bibliography or a universal performance leaderboard.
  • No measured cross-method results were generated for this library. Robustness notes summarize cited evidence and must not be read as guarantees.
  • Unverified release numbers, prices, license details and browser performance are marked explicitly rather than inferred.
  • Dataset and media rights remain separate from a code license. Face/voice/device-linkage and public hash indexes may create privacy risks even without retaining original media.
  • Text-only detection, document authorship, 3D assets, biometrics for identity decisions, moderation databases and legal ownership adjudication are outside the core image/audio/video scope.

This is a dated research snapshot. Before selecting a dependency, recheck the primary repository or official documentation, pin an immutable release/commit and model digest, resolve code/weights/dataset licenses separately, and run the evaluation protocol. This maintenance update did not activate accounts, signing services, plugins, subscriptions or public publishing.

Primary source index

All 193 sources in this edition. Review dates vary by record; this edition includes bounded updates through 2 November 2026. Links open the original websites.

  1. A leak in PRNU based source identification: Questioning fingerprint uniqueness
  2. A Self-Supervised Descriptor for Image Copy Detection
  3. AASIST ASVspoof5 baseline
  4. AASIST official implementation
  5. AcoustID commercial service
  6. AcoustID data and software license
  7. AcoustID web service documentation
  8. ACRCloud custom-file results and project types
  9. ACRCloud identification API
  10. An undetectable watermark for generative image models
  11. ASVspoof5 Phase 2 evaluation plan
  12. ATSC A/334 current standard
  13. ATSC A/335 current standard
  14. audfprint author implementation
  15. AudioSeal repository
  16. audiowmark detailed documentation
  17. audiowmark repository
  18. Automatic Detection of Internal Copy-Move Forgeries in Images
  19. AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection
  20. Bender et al., Techniques for data hiding