A bidirectional-LSTM method based on temporal features for deep fake face detection in videos

Abstract

The technology of deep fake face detection in videos is particularly important in the field of internet and finance. In order to solve the significant problem of the spread of misinformation, it is essential that deep fake face detection methods robust manipulation. This paper proposes a Bidirectional-LSTM method based on temporal features for deep fake face detection in videos. Bidirectional-LSTM method is one of deep learning methods, which has positive effect on exploiting the temporal information in videos. Facial expressions and muscle movements play an important role in an individual’s speaking style. The continuity features of temporality suffer from the working process of the fake face in videos so that deep fake detection can be used for the fake face videos. Compared with previous methods such as SVM, BP neural network, the proposed method takes less time to train the model and have a good performance on accuracy. The experimental results show that the accuracy reaches 82.65% on DFDC data set. The proposed method has an excellent effect on the robustness in detecting compressed videos with noise interference.