Research On Detection Method For DeepFake Face

With the rapid development of mobile internet and mobile computing devices,the medium for accessing and producing information has gradually shifted from text and speech to images and videos.Images and videos have gradually become the main carriers of information in the current internet.With the rapid development and maturation of deep learning technology,deepfake generation technology,which can generate false images and videos has become more easily accessible to ordinary people.Due to its ease of use,deepfake content can be mass-produced,and the misuse of deepfake technology may pose a huge threat to daily life and national security.Therefore,researching how to efficiently identify deepfake content is ignificant importance.There are two main approaches to identifying deepfake content algorithms:traditional image forensics methods and deep learning methods.The deep learning-based identification method can obtain a discrimination model by learning a large number of positive and negative samples.Compared with the traditional forensics methods,this method does not require manual design of the network,and can self-learn through a large amount of data,and still maintain good identification effects when faced with new forgery methods.This article proposes two model algorithms from two aspects: enhancing the model’s ability to focus on key areas of forgery and improving the model’s ability to extract image features,specifically:1.Discrimination algorithm based on self-attention mechanism: In order to improve the model’s attention to key areas of image content,this article proposes a discrimination algorithm SAMDIS based on a self-attention mechanism.This method uses a selfattention mechanism module to enhance the focus of the backbone network model on local details and slight abnormal information,which better guides the model to extract prominent feature representations of fake areas,and achieves accurate classification of real and fake content.On the high-quality Face Forensics++ dataset,the AUC and ACC indexes reached 99.31% and 98.99%.It also achieved the state-of-the-art level in the Celeb DF dataset,the self-built dataset SAMGAN3,and the self-built non-face dataset NFD.In the generalization experiment,the SAMDIS network had shown advantages in multiple evaluation matrix compared to the baseline network.2.Discrimination algorithm combining image texture enhancement and selfattention mechanism: By introducing the image texture enhancement module,this article proposes the SAMTDIS model,which can obtain more image texture feature information and improve the model’s ability to obtain fake traces.On the Celeb DF dataset,the ACC and AUC of the SAMTDIS network increased compared to the SAMDIS network.On the Face Forensics++ dataset,the SAMTDIS network also maintained the state-of-the-art performance,fully demonstrating the effectiveness of the image texture enhancement module.