DeepfakeDetector: video model
EfficientNet-B4 + Bi-LSTM with attention that classifies a face video as REAL or FAKE. It is the video detector of DeepfakeDetector, a diploma project.
| File | Use |
|---|---|
backbone.onnx |
EfficientNet-B4 features for 16 aligned 224ร224 face crops |
temporal.onnx |
Bi-LSTM with attention over the 16 frames, returns the FAKE logit |
last.ckpt |
PyTorch Lightning checkpoint, used only for attention weights and HiResCAM heatmaps |
Preprocessing, thresholds and the inference code are in video-detector/ of the GitHub repository. The detector
downloads the files from a pinned commit of this repository and checks their SHA-256 (video-detector/models.toml).
Training data and licence
Trained on FaceForensics++ (c23). The weights are released for research and education only, not for commercial
use, following the terms of the training data. The backbone starts from ImageNet weights from timm.
Limitations
The model sees only the largest face in 16 sampled frames. It was trained on one dataset and generalises worse to other generators and to heavily compressed or low-resolution video. Treat its output as a signal, not proof.