Utilizing Recurrent Neural Network for Deepfake Video Detection
Jui‐Feng Yeh, Ming-Jheng Shih, Jingxiang Yang, Chongyin Li · 2024
In recent years, deepfake technology has been significantly advanced, making fake videos increasingly difficult to distinguish. The face-manipulated videos have been used to interfere with political elections and impact personal reputations, threatening social and individual safety. Thus, deepfake video detection technology is essential as a countermeasure against deepfakes. Recent studies have demonstrated that spatiotemporal feature analysis is useful in identifying fake videos. Based on the result, we trained residual neural networks for facial feature extraction and recurrent neural network for temporal feature processing using three well-known datasets. We also designed a user interface to allow users to upload videos for recognition, and in response, provide a binary prediction outcome. The entire model architecture was validated with an accuracy of 85% and an F1 Score of 0.9 for predicting deepfake videos.