Research On GAN-based Facial Information Editing And Privacy Protection Methods

With the development of computer vision and the rise of live broadcasting and other industries,facial image data has become more and more important in people's daily interactions.The corresponding processing technology has also received extensive attention.This technology has various manifestations,including facial attribute information editing,style transfer,de-identification,and content restoration.In recent years,the proposal of Generative Adversarial Networks(GAN)has greatly promoted the development of generative models,and thus led to the development of facial information processing technology.Nevertheless,there are still many challenges in this field.For example,attribute information editing is difficult to deal with the accuracy and authenticity of editing results at the same time.De-identification technology also has the dilemma of insufficient diversity and poor controllability.Given these existing problems,this paper attempted to combine the efficient generation and representation capabilities of GAN with attribute information that has both stability and individual differences.And we explored the application of GAN in facial information processing technology from two directions of facial attribute editing and facial information privacy protection based on attribute editing.Overall,the research results obtained in this paper are as follows:(1)For facial attribute information editing,this paper introduced the spatial attention mechanism into the generator and used the target domain localization map generated by this module to improve the network's ability to distinguish between necessary edits and redundant edits.In addition,to improve the convergence speed of the network,this paper introduced and constructed a channel normalization algorithm based on the convolution method(CNC)based on optimizing the training strategy.During network training,the algorithm was applied in the cascaded modules and discriminator.Finally,based on GAN adversarial training,this paper constructed a pyramid discriminator and attribute classifier with parameter sharing to constrain the overall quality of the generated images.The experimental results showed that the method proposed in this paper had a better editing effect and a higher editing efficiency.(2)For facial information privacy protection,this paper further studied the impact of diverse facial attributes on facial information based on previous work and constructed a facial information privacy protection model based on attribute editing.This model continued to use the GAN architecture,which structurally retained the attribute classifier and pyramid discriminator to improve the overall quality of the editing results.In addition,the model constrained the de-identification effect with input terms and an identity-guided discriminator.Finally,this model still utilized the CNC to optimize the training process of the network.The experimental results showed that the method proposed in this paper had a better recognition feature removal effect,and the de-identification process was more controllable.