Research And Generalization Exploration Of Deep Forgery Detection Method Driven By Forgery Characteristics

With the continuous development of deep generative models and generative adversarial networks,deep synthesis technology is also gradually intelligent and precise.However,technology is a double-edged sword.Deep synthesis technology not only facilitates video production,but also brings potential risks to social development.At present,Deepfakes,which mainly focus on face tampering,have become a research hotspot in the field of synthesis.More and more forged products are circulating on the Internet,which has brought great impact on personal reputation,public opinion and even national security.Therefore,researching effective detection methods for deep forgery is an important topic to clear the network environment and maintain network security.In addition,the development of detection technology is usually lagging behind,and it has been unable to match the development speed of deep forgery.On this requirement,this thesis studies deep forgery detection methods driven by forgery characteristics and carries out an exploration of generalization detection.Firstly,this thesis proposes a deep forgery detection method based on texture difference characteristics,which finds the difference between the texture formed by the fake image and the texture of the real shot image and uses it as the main feature to classify the real and fake.Specifically,the method firstly finds the subtle texture differences of the real and fake face regions in the saliency maps of the real and fake images.Then,in order to take advantage of this difference,the method uses the saliency map as the guide map of the guided filter to amplify texture artifacts and other fake potential features.Finally,using the real and fake image with simultaneous texture amplification as a dataset to train the Resnet18 classifier to efficiently learn this texture difference,and realize the real and fake detection of face images.Secondly,this thesis proposes a deep forgery detection method based on the complementary fusion of "variability" and "invariance" of forged pixels,to discover its invariant and changed features in forged images,and use their advantages to perform authenticity classification.Specifically,this method continues the first work and continues to capture the forgery traces left by image tampering,and proposes the concept of pixel-level detail enhancement,using bilateral filtering to refine the edge of the image,and amplify the connection trace between the forged and the non-forged area;Meanwhile this method proposes the concept of pixel-level saliency invariance,and uses Quaternary Fourier Transform to map the fake image to the pixel saliency map,which proves that different forgery methods have significant invariance in different compression levels.Finally,a dual-stream network is proposed to extract features,and the two features are weighted and complemented to realize real and fake detection.Thirdly,this thesis proposes a deep forgery detection method based on the reproduction of forgery traces.The trace of the forged area is generated by the characteristic that forgery would inevitably leave fingerprints,and this trace is used as a clue for authenticity detection.Specifically,this method first proposes an end-to-end generative model to track the latent texture traces in the image generation process.Then,a self-texture attention mechanism is proposed to analyze the autocorrelation of forged regions and their long distances from non-forged regions and connect the encoder and decoder of generative model in skip connection,which can enhance the texture features in the process of image disassembly and assist the generation of forged trace maps.At the same time,this method also proposes a loss limited by the classification probability function to directly correct the generated fake traces.Finally,the generated trace map and the original image are combined and input into the Resnet18-based classifier to achieve real and fake classification.Fourthly,this thesis proposes a generalization detection method for deep forgery based on the difference data distribution,which only learns the data distribution of real videos to form an independent detection boundary,and performs generalization detection for unseen forged videos.Specifically,this method first constructs a Generative Adversarial Network(GAN)framework to learn the distribution of real face data and generate corresponding face data.Then,in order to break the high fitting state between the forged area and the background,this method proposes a generator composed of a recursive cross-attention encoder and a styleGANv2 decoder.Finally,to avoid the situation of as long as a face is normal or different faces are abnormal,this method additionally sets up a latent space encoding discriminator,and increases the ratio of latent space vector constraint loss.In order to improve the ability to detect anomalies in the latent space,the real and fake can be identified from the latent space and mapped to the generated images during model inference,so as to realize the generalization exploration of deep forgery.The research contents of the above four aspects all focus on some characteristics that must be left over from deep forgery,and are discussed in detail from two aspects of theoretical deployment and experimental verification.The experimental results obtained are also at the most advanced level in the same period.According to the research content of this thesis,the innovation points are summarized as follows: First,this thesis mines the saliency map of the fake image for its texture and uses the saliency map as a guide map for guided filtering to amplify the image texture.Second,this thesis separately visualizes the forged pixel boundaries and counts the compression invariance of different forgery methods,and proposes a comprehensive complementary twostream network to integrate the advantages of both features.Third,this thesis organically combines the generation model with forgery detection,and develops a new idea of deep forgery in the form of generating trace maps,aiming at the characteristics that the forged area would inevitably produce traces that are different from the nonforgery area.Fourth,in this thesis,Generative Adversarial Network(GAN)is used in data reconstruction of deep forgery,and the generalization discrimination of deep forgery is explored in view of the difference between the potential distribution of forged data and real data.