Research On Face Compression Image Of Deepfake Detection

Deepfake technology has attracted considerable attention in recent years due to its extraordinary realism that makes it difficult for the human eye to distinguish.However,due to the loss of forgery traces caused by compression,detection of compressed forged images has been a difficult task.Although some work has been done for compressed forgery detection,it suffers from two shortcomings.The first is that these works tend to ignore the robustness of their models or the backbone models they rely on and lack the ability to detect other compressed forgeries.The second point is that two-stream structure is used by some current works consisting of two different branches to learn HD and compressed image information separately,which has the disadvantage of requiring prior knowledge of the quality a priori information of the test data during the testing process.To address the above two problems,the following work is carried out in this thesis:First of all,for the first point,this thesis proposes a method to increase the detection performance of existing models for other compressed images.By analyzing the influence and reason of the compressed forged traces on the feature layer of the detection model,the problem and theoretical analysis combined with the adversarial attack are established.After establishing the relationship between the feature layer and the input,a new model structure and training method are designed.In the process,two loss functions and a feature balance module are also designed to make the model better learn the difference of the feature space of distribution representation.In addition,for the second problem,the causal effect problem is introduced by this thesis into the compressed image forgery detection task.It continues to improve the original model based on the established problem by reorganizing the previous mixed inputs into non-strict pairs of inputs,and introducing a new loss function in the feature extraction part to reduce the impact of compression as causal invariance on the detection task.In addition,a new model framework is designed to combine the causal effects with the traditional depth falsification detection model,and a 1x1 convolution module is introduced to improve the model structure to better utilize the information in the spatial domain of the image.Finally,this thesis conducts a comprehensive experiment to verify the first method under 5 kinds of backbone networks and 4 kinds of data distribution,and achieved good performance in terms of versatility,adaptability and accuracy improvement,and the average accuracy rate increased by about 6%.For the second method,this thesis compares all the data with the existing advanced methods,and the results show that the method proposed in this thesis has advantages in all four datasets.For the completeness of the experiment,this thesis fully carries out the ablation experiment and visualized it to increase the interpretability,and verified the effectiveness of the method proposed in this thesis from various perspectives.Several improvements have been made for the task of deepfake detection on compressed images.