dessa-oss/DeepFake-Detection

Fine-tuned ResNet18 backbone trained on combined FaceForensics++ and real-world YouTube datasets to address generalization failures in existing deepfake detectors. Orchestrates large-scale hyperparameter search and dataset experimentation through Dessa's Atlas scheduler, with preprocessed frame extraction via FFmpeg and caching mechanisms for efficient training. Demonstrates that detection models require continuous real-world data augmentation, moving beyond synthetic benchmark datasets to achieve practical robustness across multiple manipulation techniques.

643 stars. No commits in the last 6 months.