GRACE deepfake detection models
Trained weights for GRACE (Generalisable, Robust And Concept-Explainable detection), capstone project PW26_RJ_02, Department of CSE, PES University (Aug–Dec 2026).
- Code, results and report: https://github.com/pes1ug23cs178/CAPSTONE_PW_26_RJ_02
- Live demo: https://huggingface.co/spaces/grace-cap/grace-deepfake-check
Files
| File | What it is |
|---|---|
grace_teacher.pth |
Final teacher (F1): trainable parts only (LoRA adapters on CLIP ViT-L/14, DCT branch, gated fusion, 10-clue concept head and decision layer). The frozen CLIP backbone is downloaded through timm when the model is built. |
grace_teacher_config.yaml, _temperature.json, _val_threshold.json |
Model config, calibration temperature (fitted on FF++ validation), decision threshold |
grace_student.onnx + grace_student.json |
4.7 M-parameter EfficientNet-B0 student (S3) for CPU inference, plus its concept weights |
grace_student_s3.pth and its config / temperature / threshold |
The same student in PyTorch |
SHA256SUMS |
Checksums |
How it decides
Each aligned 224×224 face is scored on ten named forensic concepts (blend strength, colour mismatch, blur, over-sharpening, resampling, mouth edited, eyes edited, whole face swapped, misaligned geometry, irregular boundary). The fake logit is a weighted sum of these concept scores plus a small penalised residual, so the explanation is part of the computation. Concept labels come from the recipe of self-blended training forgeries.
Training and evaluation
- Trained on FaceForensics++ (c23; Deepfakes, Face2Face, FaceSwap) with self-blended images. DFD and Celeb-DF v2 are never used for training.
- Video-level AUC (three seeds) before the switch to the official Celeb-DF list: FF++ 0.997, DFD 0.951, Celeb-DF 0.901 (interim 518-video set). Results on the official Celeb-DF test list are in the GitHub repository.
Limitations
Visual, frame-based detection of face swaps. It can be fooled by new generators (for example diffusion-based faces), very strong video compression (H.264 CRF 40) and extreme head poses. Use the output as evidence, not proof. The datasets used for training are not redistributed.