Research On Face Forgery Image Detection Based On Deep Learning

Face forgery technology refers to the use of generative models for face identity and expression tampering,the most representative of which is "deepfakes".At present,a large number of malicious videos generated by such algorithms have been spread in the network,posing a serious threat to personal privacy and social stability.Therefore,face forgery detection has become an urgent problem to be solved.To address the problems of low detection accuracy and poor interpretability of existing deep learning-based face forgery detection,this thesis designs two detection schemes for samples with known forgery means from the perspective of frequency domain features and detail features,respectively,and achieves ideal detection results.To address the problem that existing detection schemes will inevitably make the model tend to over-fit the traces and distortions in a particular face manipulation method,and the performance in unknown data distribution shows serious degradation,i.e.,insufficient generalization,this thesis proposes a detection scheme with high generalization from the perspective of exploring forgery traces with generalizability.Specifically,it includes:(1)From the perspective of frequency domain features,this thesis proposes a face forgery detection scheme oriented to the high-frequency component of shallow features.Existing detection methods based on high-frequency features are usually based directly on the original image,and the performance of such algorithms is poor if the image artifact features are not obvious.In this scheme,extraction,enhancement modules and image gradient loss are designed separately in order to enhance the learning of high-frequency components of such shallow features.The experimental results demonstrate that this scheme can effectively improve the accuracy of the detection model and reduce the number of parameters and computational effort of the model by learning the high-frequency components of such shallow features;(2)From the perspective of fine-grained features,this thesis proposes a face forgery detection scheme based on fine-grained features of images.This scheme defines the face forgery detection task as "locating highly distinguished regions and extracting fine-grained features".Two specially designed feature extraction modules are used to obtain region-level and fine-grained features,and a multi-branch network with shared parameters is used for learning.(b)The experimental results demonstrate that this scheme achieves desirable detection results in a known data distribution using multi-level features of images;(3)From the perspective of detection model generalizability,this thesis proposes a face forgery detection scheme based on forgery cue mining.This scheme proposes that for the face forgery detection task,local cues play a more critical role than global semantics.Therefore,exploring the prevalent local forgery traces is the key to the face forgery detection task.This scheme proposes two functional modules to reveal and localize deeper forgery features,and employs a sensitive patch branch shared with the main network parameters to locate important subtle forgery traces.Experimental results show that this detection scheme for generalizable forgery trace mining has better performance both for unknown data distributions and highly compressed images.