Next generation of verification.
Synthetic media is cheap to make and hard to spot. The answer is not a score. It is evidence a person can act on.
We fuse forensic signals across media, identity, and behavior so a reviewer can see why something is flagged, not just that it was.
Learn more about usDetection you can measure.
How the numbers are made, from the public technical papers. Read the papers ↗
- backbone
- multi-scale CNN, attention-augmented blocks, linear cost
- parallel pathway
- 8×8 DCT spectral residual
- context head
- provenance · metadata · benign-edit state
- fusion
- p = σ( Σᵢ wᵢ·ℓᵢ + b ), calibrated
- abstain
- |p − 0.5| < τ → review
- parameters
- 8–15 M (ViT-Base/Large: 80–300 M)
- model size
- 90.24 MB
- latency
- 14.07 ms NPU · 37.22 ms GPU · 74.13 ms CPU
- memory
- 114.35 MB FP16 · 137.69 MB FP32
- NPU speed-up
- 5.27× over CPU
GaussMass is a family of homegrown computer-vision models. An image enters a hierarchical convolutional stem that extracts features at four scales: early layers respond to compression seams, resampling and blending boundaries; deeper layers respond to lighting, shadow and facial-geometry inconsistencies. Lightweight attention is inserted inside the convolutional blocks and biased toward artifact-prone regions (eyes, mouth, face boundary), so cost stays linear in image size rather than quadratic in patch count.
In parallel, a discrete cosine transform over 8×8 blocks produces a spectral residual that exposes the periodic signatures upsampling and diffusion pipelines leave behind. A context head reads provenance (content credentials when present and valid), file metadata and a benign-edit state, so tone grading, portrait mode or social recompression are treated as context rather than as evidence of synthesis. Missing provenance is treated as unknown, never as authenticity.
The evidence families are combined as a weighted sum of log-odds, temperature-calibrated so that a reported 0.9 means 0.9. If the margin from 0.5 is below the abstention threshold, or the pathways disagree, the system returns review-required instead of a verdict. The output is decision support with the evidence attached: verdict, likelihood, evidence categories, media hash and timestamps.
One platform, three surfaces.
- Deepfake detection
- Image, video, audio, metadata, spectral, temporal, and web-intelligence signals fused into a clear verdict, with the evidence behind it.
- Fraud defense
- Escalate suspicious media, impersonation attempts, and account-level risk into cases your team can resolve.
- Proof-of-humanhood
- Verification workflows, liveness status, and audit trails for marketplaces, platforms, and trust teams.
Tech Blog.
- Deepfake fraud, by the numbersThe figures that hold up, where they come from, and what they do not say.
- Four incidents that defined deepfake fraudWhat the public record shows about how synthetic media attacks actually unfold, and what each one taught.
- What the law now says about synthetic mediaFour rules in force or arriving in the United States, the European Union and the United Kingdom, and what each one asks of platforms and tools.
See what your team is missing.
