BMNet: Enhancing Deepfake Detection Through BiLSTM and Multi-Head Self-Attention Mechanism

When forgery techniques can generate highly realistic videos, traditional convolutional neural network (CNN)-based detection models often struggle to capture subtle forgery features and temporal dependencies. Most existing models focus on feature extraction from static frames, neglecting the temporal correlation in videos, which decreases accuracy in detecting dynamic forged videos. Furthermore, the ability to detect localized forgery features remains insufficient. To address these limitations, we propose a deep forgery detection framework named BiLSTM Multi-Head Self-Attention Network (BMNet). By leveraging the Bi-Directional Long Short-Term Memory Network (BiLSTM) for modeling temporal dependencies between video frames, and the Multi-Head Self-Attention Mechanism (MHSA) for capturing features from different regions of an image, BMNet more effectively identifies dynamic and local forgery features. In our experiments, we extract features from 68 facial landmarks of each video frame and evaluate detection performance on four datasets: UADFV, FF++, Celeb-DF and DFDC. The results show significant improvements over traditional methods. We further validate the necessity of each network component through ablation studies, demonstrating that BMNet achieves accuracies of 95.54%, 92.18%, 80.20% and 84.72% on the FF++, UADFV, Celeb-DF and DFDC datasets, respectively, indicating its superior performance in deep forgery detection.