Research On Forensics Of Face Tampering Based On Deep Learning

In the context of the development of Internet applications and information security technology,deep forged image generation technology has developed rapidly.It also brings risks and hidden dangers to network information security.The Internet is full of many "face-changing" images or images based on deep learning.If these "deep face-changing" pictures or videos are maliciously manipulated,they will greatly affect the credibility of network information.Therefore,how to accurately detect forged images is an important research area of biometric security.The main results of this article include:1.According to the current mainstream open-source methods,all of the features that can be distinguished between true and false are automatically learned through the CNN model.The characteristics of forged features are rarely considered,to design a detection model more suitable for deep forged image detection.Our method is different from the conventional CNN classification and recognition network,and a detection network that meets the characteristics of forged traces is designed.First,the Resnet34 convolutional neural network is used to extract the preliminary forgery features of the image to be detected.Then the receptive field of the network model is improved through the hole convolution to better capture the scattered forgery traces,and the spatial attention module performs the weighting of the weights.To reduce the interference of other background information,the multi-layer fully connected neural network is used to classify the features to accurately classify the deep forged images.Experiments are conducted on the three mainstream data sets of Faceforensics++,Celeb-DF,and DFDC.Our method achieves better results than other current methods.2.Propose a multi-class detection network for deep forgery of face.Most of the current deep forgery forensics algorithms are a two-category structure,and there is a lack of traceability to the forgery method.The depth of face forgery and tampering area is mainly concentrated on the face,in order to eliminate the interference of irrelevant information,the method uses a dual-stream network to fine-grained classification of the deep fake method of human faces.One of the streams is based on channel attention,used to focus on key network channels.The other flow is based on spatial attention,used to extract key regions on a single feature map.The method uses the attention mechanism to help the network learn the main features better,avoiding the interference of other secondary information on feature decision-making.Through experiments on mainstream face depth forgery datasets,better results than current mainstream detection methods have been achieved.