FAME is a lightweight, spatio-temporal deep learning architecture designed for Deepfake model attribution — identifying the source generative model behind manipulated face-swap videos. It combines VGG-based frame-level representation with LSTM-based temporal attention for robust multi-class classification.
🔒 Note: The source code and pretrained models will be released shortly upon acceptance/publication.
- 🧩 Combines truncated VGG19 for spatial feature extraction with Bi-LSTM for temporal modeling
- 🔍 Attention module to highlight salient frames in video sequences
- 💡 Designed for attribution across multiple Deepfake generation models, not just detection
- ⚡ Lightweight and efficient, suitable for real-time or resource-constrained deployment
FAME has been evaluated on multiple public benchmarks:
| Dataset | Accuracy (%) | Macro F1 Score | AUC |
|---|---|---|---|
| DFDM | 97.3 | 0.975 | 0.999 |
| FaceForensics++ | 96.4 | 0.968 | 0.997 |
| FakeAVCeleb | 84.6 | 0.841 | 0.986 |
FAME shows strong generalization across different types of face-swap and GAN-based Deepfake generators.
The codebase will include:
- Full PyTorch implementation of the FAME model
- Training & evaluation scripts
- Support for DFDM, FF++, and FakeAVCeleb datasets
- Pretrained model checkpoints (to be released)
⏳ Coming Soon: The code and models will be publicly released on this repository upon official acceptance/publication.
If you find our work useful, please consider citing:
@article{AHMAD2025128571,
title = {FAME: A Lightweight Spatio-Temporal Network for Model Attribution of Face-Swap Deepfakes},
journal = {Expert Systems with Applications},
pages = {128571},
year = {2025},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.128571},
url = {https://www.sciencedirect.com/science/article/pii/S0957417425021906},
author = {Wasim Ahmad and Yan-Tsung Peng and Yuan-Hao Chang},
keywords = {Face-swap Deepfakes, Deepfake Model Attribution, Attention Mechanism, Multimedia Forensics, Information Security}
}