Research and Performance Evaluation of Deepfake Video Detection Method Integrated Spatio-Temporal Attention Mechanism

With the rapid development of deep learning and computer vision technology, Deepfake technology has been able to generate highly realistic fake videos, which brings serious security and ethical challenges to society. Existing detection methods have insufficient ability to deal with the evolving deepfake technology, especially when dealing with the spatio-temporal characteristics of videos. To solve this problem, a deepfake video detection algorithm based on Spatio-Temporal Attention mechanism is proposed. By fusing the Spatio-Temporal information of the video, an Attention-Driven LSTM network (ADLNet) is designed, which uses the attention mechanism to accurately identify the forgery traces and significantly improves the detection effect. The ACC and AUC on the public deepfake dataset FaceForensics++ are 97.17%and 99.33%respectively, and the ACC and AUC on the Celeb-DF dataset are 96.78%and 98.36%respectively. The experimental results and visualization verify that the proposed ADLNet model has strong robustness and application prospects in extracting spatio-temporal artifacts.