STANet: Spatio-Temporal Attention Network for Video Forgery Detection

In today's digital era, detecting forged video content has become increasingly important for maintaining trust and authenticity. This paper proposes a hybrid deep learning-based Spatio-Temporal Attention Network (STANet) for video forgery detection, capable of identifying forgery in videos with both static and dynamic backgrounds. The proposed STANet utilizes Convolutional neural network (CNN)-based InceptionV3 to extract spatial features from each video frame; then, the spatial features are processed through Bi-directional Long short-term memory (Bi-LSTM) to capture temporal dependencies across frames, allowing the network to differentiate between real and forged frame sequences. The key feature of the proposed method is the attention mechanism that highlights important frames within the sequence and assigns higher weights to the frames that are more relevant for determining whether a given video is real or forged. We evaluate the performance of STANet on a video forgery dataset consisting of videos captured with 5 camera devices. The experimental results show that the STANet outperforms existing methods, achieving high detection accuracy and robustness against different forgery techniques.