Research And Implementation Of Deepfake Detection Technology Based On Swin Transformer

Deepfake is a portmanteau of Deep Learning and Fake.This technology allows existing images or videos to be superimposed onto target images or videos.The abuse of this technology has increased the complexity of information dissemination on the internet,with criminals using it for defamation,fraud,extortion,endangering national security,and harming individual and public interests.Therefore,effective detection of these fake contents has become an urgent issue to be addressed.Based on multi-channel information extraction methods,hybrid scaling methods,and dual-stream attention methods,this article proposes two deepfake detection methods and further implements a convenient and efficient deepfake detection system.The main research content of this article is as follows:(1)This thesis proposes a deepfake detection method based on a multi-channel Swin Transformer.By extracting the channel information in the color space,facial feature space,and frequency domain space of the image and stacking it as multi-channel information,multi-channel information extraction of the data is achieved.These multi-channel pieces of information are input into the Swin Transformer feature extractor for detection,and the model’s backpropagation process is performed through the cross-entropy loss function and Adam W optimizer,ultimately obtaining the detection model.After testing on the public dataset Face Forensics++(FF++),the experimental results show that the deepfake detection accuracy of this method reaches 94.71%,which is superior to other detection methods.(2)This thesis proposes a deepfake detection method based on a compound scaled dual-stream attention network.This method detects deepfake videos by combining compound scaling modules and dual-stream attention modules based on Swin Transformer.The compound scaling module consists of residual downsampling,fusion convolution,and compression convolution,achieving more efficient local feature extraction.The dual-stream attention module,by combining selfattention mechanisms and channel attention mechanisms,achieves feature extraction in both the global dimension and the channel dimension.In the overall architecture design,the compound scaling module is responsible for extracting shallow local features of the data,while the dualstream attention module is responsible for extracting deep global features of the data.Experiments on the FF++ dataset show that the deepfake detection accuracy of this method reaches 95.62%,proving its superiority.(3)In this thesis,a deepfakes detection system is designed and implemented.The system adopts a standard browser/server architecture,with the front end built on the Vue framework and the back end developed using the Flask framework.For asynchronous tasks,the system uses Redis as the message queue,with Celery responsible for processing tasks in the queue.In addition to providing basic user rights and record management services,the aforementioned deepfake detection methods are embedded separately into the system as modules,providing users with convenient deepfake detection services.