Research On Differential Privacy Protection Techniques For Image Data Based On Generative Adversarial Networks

With the popularity of mobile smart devices and the rapid rise of emerging technology fields such as the Internet of Things and Big Data,the amount of data has exploded.Data mining and analysis using rich data can not only promote the development of the field of science and technology but also become a new engine to drive economic development.Among them,the analysis and mining technology for image data is becoming more and more mature,and great achievements have been made in both theory and practical application.However,the resulting privacy leakage problem has become more and more serious.Most of the existing methods for image data privacy protection are based on image compression or anonymization techniques.Although these two methods solve the problem of private information leakage to a certain extent,they destroy the features and clarity of image data,and the anonymization technique needs to make assumptions about the attacker's background knowledge in advance,which is difficult to do in the context of big data.Differential privacy protection technology can provide privacy protection for attackers with arbitrary background knowledge,this article combines this technology with a generative adversarial network and proposes the following two privacy protection frameworks:(1)dp-WGAN image data privacy protection framework.First,the framework uses differential privacy protection techniques combined with generative adversarial networks to train a generative model with differential privacy and uses this generative model to generate synthetic data to complete the analysis task instead of sensitive data,which ensures that privacy of sensitive data is not leaked.At the same time,a series of optimization strategies are proposed during the model training process to accelerate the model convergence and improve the model accuracy and the utility of synthetic data.Finally,the model is evaluated using a real large-scale image dataset to verify the effectiveness of the dp-WGAN privacy protection techniques framework.The experiments demonstrate that the dp-WGAN privacy protection techniques framework provides theoretically guaranteed privacy protection for image data in addition to retaining the desirable utility of sensitive data.(2)RDP-WGAN image data privacy protection framework.To improve the effectiveness of the privacy protection framework,this paper proposes an RDP-WGAN privacy protection framework based on the dp-WGAN framework.In this framework,0)9)? 4)differential privacy protection technology is deployed in the training process of generative adversarial networks,which provides stricter privacy protection for model and sensitive data.At the same time,a series of optimization strategies are used to further improve the training process of the model,reduce the amount of noise,improve the accuracy of the model,and make the synthetic data more practical.Finally,the training and evaluation on real image data conclude that the RDP-WGAN privacy protection framework significantly improves the utility of the dp-WGAN framework,giving the framework greater privacy protection capabilities and enhancing the utility of synthetic data.