Research On Detection Technology For Deepfake With Different Quality

The emergence and development of Deepfake technology makes it impossible to guarantee the authenticity of video content.At the same time,the rapid spread and wide attention of Deepfake content makes the negative impact and potential threats more and more serious.Therefore,the research on Deepfake detection technology is of great significance to the protection of personal privacy and the maintenance of social trust system.The research content of this thesis is the detection for Deepfake with different quality.Existing detection methods have a significant drop in detection accuracy when facing compressed Deepfake images and videos.In this thesis,the detection of Deepfake images and videos with different degrees of compression is defined as Deepfake detection with different quality,and two detection methods for Deepfake with different quality are proposed.The main contents are as follows:(1)A Deepfake detection method based on frequency-domain filtered residual map is proposed.The post-processing operation of Deepfake technology is mainly to erase the forgery traces in the RGB domain of images and videos,and the forgery traces in the frequency domain are relatively less affected.Therefore,this thesis proposes to combine the features of the RGB domain and the mid-high frequency domain of the images to mine the forgery traces.Firstly,the image is subjected to Haar wavelet transform,and then the grayscale image of the original image and the low-frequency information map are subjected to residual operation to obtain the residual map of the medium and high frequency information of the image.Finally,the original image and the residual map are concatenated and input into the convolutional neural network for authenticity classification.This method can extract the differential features of real and fake images in RGB domain and frequency domain at the same time,and mine richer forgery traces,thereby reducing the impact of the compression operation on the detection performance and improving the detection performance for Deepfake with different quality.(2)A Deepfake detection method based on feature similarity between face region and background is proposed.Most of the existing methods extract domain-specific features of images or videos to detect Deepfake,which are easily affected by compression operations and targeted by forgery techniques.This thesis proposes to detect Deepfake images by exploiting the inherent difference in the feature similarity between face region and background of real and fake images.Since the steps in the imaging process of each region of the real face images and the videos are the same,the noise feature distribution of each region is consistent and continuous.However,the manipulation of the face region in the fake images will destroy this characteristic,and this is an unavoidable problem of Deepfake technology.Therefore,the difference in feature similarity between different regions of the real and fake images is difficult to eliminate in the compression operation,which ensures the robustness of the method to compression operations,thereby improving the detection performance for Deepfake with different quality.(3)This thesis conducts sufficient experiments to evaluate the proposed two detection methods in uncompressed,lightly compressed and heavily compressed scene respectively,and compares them with various methods in the field.The average detection performance of the proposed methods on Deepfake with different quality is then evaluated based on the experimental results of the three scenarios.The experimental results show that the two methods proposed in this thesis achieve better detection performance than the benchmark and comparative methods on Deepfake with different quality,and the method based on the feature similarity between the face region and the background achieves the best performance in the field on the three sub-datasets in the heavily compressed scene.