Forged Facial Video Detection Framework Based on Multi-region Temporal Relationship Feature

Face generation and manipulation techniques based on deep learning has enabled the creation of sophisticated forged facial video, which is indistinguishable by human eyes. However, illegal use of deep fake technology will have a serious impact on social stability, personal reputation and even national security. Therefore, detection technology of fake facial video is of great significance to protect national security and maintain social order. Although the existing video-based fake face video detection technology has achieved good detection performance on the public fake face video database, there are still the following problems: (1) the existing technology uses 2D attention mechanism to obtain local region features from face images, and lacks a 3D attention mechanism to obtain local area features from face videos; (2) After obtaining local area features, the existing technology is directly used to classify or only model the inter-regional relationship of images, without modeling the temporal relationship between regions of the video. This paper proposes a fake facial video detection framework based on multi-region temporal relationship features, including designing a three-dimensional attention mechanism to extract local features of multiple regions of the face from the video. In order to model the time series relationship between different facial areas, a time series graph convolution neural network is also introduced to extract the time series relationship features between multiple areas. In order to model the time sequence relationship between different face regions, the convolution neural network of time the sequence diagram is also introduced to extract the characteristics of the time sequence relationship between multiple regions. Through the change characteristics of time sequence relationship between facial regions, the timing inconsistency of face video is detected, so as to determine whether the face has been deeply forged. Through experiments on multiple data sets, the experimental results of the model accuracy test show that the method proposed in the present invention achieves the highest detection accuracy, and the accuracy in the FF++ (LQ) data set is 18.19% higher than the benchmark method. The experimental results of the generalization ability test show that the method proposed in the present invention achieves the highest generalization performance, and the detection accuracy of the Celeb-DF data set is 11.92% higher than that of the benchmark method.