DeepTrace: Explainable Deepfake Video Detection using BiLSTM and Dual Attention Mechanisms

Abstract

The rapid growth of deepfake videos has become a significant threat to the believability of visual media, enabling misinformation, manipulation, and damage on a large scale. Even though recent detectors using deep learning methods exhibit high accuracy in identifying deepfake videos, these detectors are often black boxes that provide little insight into the detection process. Enhancing explainability is therefore a critical requirement for reliable deepfake detection. The proposed system consists of ResNet50 convolutional backbone for extracting spatial features and a two-layer BiLSTM for capturing temporal features through video frames. In order to improve the efficiency of the model and increase the interpretability of its predictions, DeepTrace makes use of two attention mechanisms-temporal and channel attention-that focus on important video frames and facial features, respectively. The Grad-CAM approach is also used in the model to produce visual explanations based on heatmap generation. The training of the model was done on Celeb-DF while the evaluation was done in domain and cross-domain. While the DeepTrace performs extremely well in-domain for Celeb-DF, the cross-domain performance on FaceForensics++ is lower. The results show that DeepTrace provides competitive detection performance along with high levels of interpretability.