Exposing DeepFake Videos Based On Interframe Features

Over the past few years,with the advent of Deep Fake videos,facial data manipulation has become a serious threat.In a Deep Fake video,the facial attributes of a person can be easily replaced.These videos are so authentic that traces of manipulation are hard to be detected.The abuse of these manipulated videos has a serious political,economical and social effect.All these effects make the detection of Deep Fake videos an important and challenging task in network security defense.In spite of the majority of efforts made around the detection of interframe features,the incoherence of facial data crossing frames is often ignored by existing detection methods.These methods can be easily countered through optimized facial synthesizing techniques,resulting in performance degradation.Thus,in this paper,we propose a novel deep fake detection method,which captures the sharp changes in terms of the facial features caused by the composite video.We take the Gated Recurrent Unit,together with the self-attention mechanism and convolutional neural network to build a fast and efficient Deep Fake video detection system,which can make full use of the temporal feature across the Deep Fake video frames and achieve the goal of accurate detection.Since the spatial information of video frames also plays a very important role in DeepFake video detection,we utilize the Convolutional Long Short-term Memory(Conv LSTM)to enforce both spatial and temporal information of Dee Fake videos.Meanwhile,we apply the attention mechanism to emphasize the specific facial area of each video frames.We design a decoder to further fusion multiple frames information for more accurate detection results.Compared with the original detection algorithm based on the Gated Recurrent Unit,the algorithm performance has been greatly improved.In order to verify the method proposed in this thesis,we evaluated the proposed method on the Face Forensics++ dataset.Experimental results and comparisons with the state-ofthe-art methods demonstrate that our framework achieves superior performance.