Deepfake detection for images and video

sealverity.ai checks photos and video for manipulation and AI generation by running several independent detectors at once and combining their results into one explainable score. Investigators see not just a number, but where in the file the problem is likely to be and how strongly the detectors agree.

Why one detector is not enough

Every deepfake detector is trained on particular generators and compression conditions. A single model can be confident and wrong when it meets content it was not trained on. sealverity.ai sends each file to all live detectors your organisation has enabled, in parallel, and combines their scores using weights you set per media type.

Each weight change requires a written reason and is written to the audit log, so the way a score was produced can always be reconstructed later. A detector with weight zero is not called and cannot affect the result.

Confidence bands and disagreement

The combined score comes with an agreement band. When detectors broadly agree the band is high; when they disagree strongly the result is marked inconclusive and the item is sent to the review queue instead of being presented as a verdict. If only one detector returned a score, the result says so (for example, single detector, 1 of 4), and if none returned a score the analysis shows no detector result rather than a guess.

See where manipulation is likely

For images, heatmap overlays highlight suspicious regions, with a legend and a selector to view each detector's map. Where a detector does not return regions, face boxes can be used as a fallback signal.

For video, frames are sampled at one frame per second and scored, producing a timeline and a list of the five most suspicious timestamps so analysts can jump straight to the moments that matter. These overlays and timelines are stored with the analysis and reproduced in reports.

Image heatmaps with per-detector selection
Video 1 fps timeline with top suspicious moments
Metadata and EXIF extraction at intake
C2PA content credentials read automatically when present
Reverse image search to find the earliest known appearance

Measure accuracy on your own data

Detection quality varies by content type, so sealverity.ai includes a benchmark module. Administrators upload labelled samples into private benchmark sets, run every detector and the ensemble against them, and review per-detector metrics, threshold sweeps and generator breakdowns. Recommended threshold changes are audited, and drift alerts warn when results change over time.

Humans decide

Detection scores are probabilities, not proof. sealverity.ai never makes the final decision: a trained analyst does, with a recorded reason, and overrides require approval. Every action is kept in a hash-chained, append-only chain of custody, and findings can be exported as a signed forensic report.

Part of one investigation platform

This capability is part of sealverity.ai, a single platform for media authenticity and investigations. Findings from images, video, audio, live calls, identity checks and social and news monitoring can be linked to cases, entities and a link graph, with risk scores that show which factors contributed. Access is protected by mandatory multi-factor authentication, single sign-on and SCIM provisioning, five roles, retention policies and legal hold. Organisations can bring their own self-hosted detectors or switch on sovereign mode, which blocks every external call on the server and logs each blocked attempt. Integrations include a browser extension, Slack and Teams, email scanning, webhooks, SIEM connectors and a REST API.