Face Forgery Detection Based On Deep Learning

With the development of technologies such as deep learning and Generative Adversarial Networks(GAN),it has become easier to manipulate faces,and the forgeries generated by such new technologies are often able to deceive the human eye.There are already a variety of wellknown face forgery technologies that are publicly available,which may have many negative impacts on society.These technologies may be used to fabricate fake news,engage in fraudulent activities,etc.,which may have serious challenges to personal privacy,social stability and national security.At present,there are some detection technologies for face forgeries,which generally focus on specific forgery traces and limited local features.These methods are difficult to apply in real scenes,they cannot fully capture the steganographic features of face forgery,and are susceptible to noise,video compression and other factors.Therefore,it is necessary to study a high-precision detection algorithm suitable for complex and diverse scenes with high robustness.Face forgery will inevitably leave forgery traces,which are shallow representations and are difficult to be discovered by ordinary convolutional networks that focus on high-level semantic information.Therefore,the model needs to be able to analyze shallow semantics,focus on locally tampered areas,exclude redundant features,and avoid the effects of noise and video compression.Based on the information above,firstly,a face forgery detection framework based on Convolutional Neural Network is constructed,which can realize end-to-end forgery detection after sampling and preprocessing the original data.By introducing a One-Shot Aggregation(OSA)module,the low-level pixel features and high-level semantic features are aggregated to avoid the erase of forgery information.The One-Shot Aggregation module and global average pooling,fully connected layer,Sigmoid activation function and other modules are integrated into the detection framework to form the Vov Net model,which achieves high accuracy.Secondly,since it is difficult for convolutional networks to capture local tampering features,a channel attention mechanism and a spatial attention mechanism are introduced,which are combined with OSA module,so that the model can adaptively predict potential key regions and local features.During training,the triplet loss and the cross-entropy loss are weighted and summed to help the model better learn the difference between real and fake samples.A large number of experiments show that the face forgery detection model based on channel spatial attention and Vov Net achieves the leading performance on the Face Forensics++and Celeb-DF datasets,and also has good generalization and anti-interference ability.