DeepGuard
Frame-level deepfake detection
A video is sampled into evenly spaced frames, each face is cut out in the browser, and an EfficientNet-B0 scores every crop on the server. You get one probability per frame instead of a single vague verdict.
Verdict
—
Fake probability
—
Per-frame score
No face was found in any sampled frame, so full frames were scored instead. This result is not trustworthy.
About this model
EfficientNet-B0, 4.05M parameters, trained in two stages (frozen backbone, then the last two blocks fine-tuned) on face crops from the DFDC subset.
| Test accuracy | 51.3% |
|---|---|
| Test ROC-AUC | 0.548 |
| Test F1 (fake) | 0.330 |
Honest numbers: the current checkpoint is only marginally better than chance on held-out data. Treat every output as a research demo, not as evidence about a video.