The project proposes a hybrid deep learning architecture for detecting deepfake videos by combining:
- EfficientNetB0 for lightweight and efficient spatial feature extraction
- Gated Recurrent Units (GRU) for modeling temporal dependencies between video frames
The model effectively identifies subtle spatial-temporal inconsistencies found in manipulated content.
- Multi-scale spatial feature extraction using pre-trained EfficientNetB0
- Sequential learning via GRU to detect frame-level inconsistencies
- Lightweight design suitable for real-time and low-resource environments
- Achieved 89.71% accuracy and 94.56% AUC-ROC on benchmark datasets
| Metric | Score |
|---|---|
| Accuracy | 89.71% |
| Precision | 88.18% |
| Recall | 82.11% |
| F1-Score | 85.04% |
| AUC-ROC | 94.56% |
Tested on combined datasets: Celeb-DF (v2) and DFDC
See model_architecture.png for visual representation.
- Celeb-DF (v2): Celeb-DF
- DFDC: Deepfake Detection Challenge Dataset
Note: Due to licensing restrictions, datasets are not included in this repo. Please download from official sources.