Research And Implementation Of Generative Adversarial Network Based Image Privacy Protection Algorithm

With the rapid development of the Internet,a growing number of people become dependent on the network.The privacy information is directly uploaded to the network and is illegally collected and used by enterprises or individuals,thus causing the privacy of Internet users to be leaked and violated.What's more,the applications of computer vision technologies in social networks has raised concerns about privacy leaks.For example,the real-world developments of face recognition on mobile phones may lead to the disclosure of private information such as personal identity.Facing privacy leaks,individuals need effective methods to protect privacy.To solve the above problems,this paper first propose a privacy protection algorithm based on generative adversarial network.Combined with adversarial examples technologies,our method could protect privacy without destroying the image utility,so that after protected,these images could still be uploaded to the social networks.Existing privacy protection methods mainly change sensitive areas of the image,such as facial and text,to protect privacy,or to remove objects that involve privacy from images or video.However,these techniques struggle in tackling problems for which images contain useful in-formation.Specifically,while sensitive information is removed,the necessary information that enables images to be correctly identified by friends may also be removed.The undesired side effect of hiding sensitive information results in the situation where the image utility is lost.The main goal of this paper is to solve the problem that how to protect the privacy of photos without destroying useful information automatically.Take this into account,we harness adversarial images to address the trade-off between privacy and the image utility.Recent work has shown that Deep neural networks(DNNs)are vulnerable to adversarial image.A well-trained image classifier based on DNNs could be confused by images with small perturbation.However,rarely exploited for privacy protection,the existing adversarial image algorithms mainly use the gradient information to guide image generation and rely on the gradient calculation of the cost function concerning the input image.To solve these problems,this paper proposes a generative feed-forward network-based approach which takes original images as inputs and outputs privacy images.The generated privacy images can reduce the accuracy of the neural network hosted by the service provider so that privacy can be protected.From the human vision,these privacy images look almost same as the original images.The experiment shows that the privacy images are misclassified which indicate that our approach can prevent the privacy from leakage with considerable improvement.At the same time,this paper also takes into account the replacement of the privacy areas in images to protect the privacy information.Unlike the first approach,in some cases,the privacy areas in images need to be processed directly,such as the background information.For these areas,not only the identification network on the social network needs to be deceived,but also the human vision needs to be deceived.In view of this situation,this paper also proposed a real-time privacy area replacement system,which directly partitioned the privacy area in images to improve the user experience.