Collaborative Face Privacy Protection Method Based on Adversarial Examples in Social Networks

The face image de-identification method commonly used to protect personal privacy on social networks faces the problem of not being able to assure the usability of image sharing. Although the privacy protection method based on adversarial examples meets this requirement, there are still potential hidden problems of information leakage and the compression processing of images by social networks will weaken the privacy protection effect. In this paper, we propose a collaborative privacy protection method based on adversarial examples for photo sharing services on social networks, called CSP3Adv(Collaborative Social Platform Privacy Protection Adversarial Example). We use the perturbation transfer module, which avoids the information leakage caused by accessing to the original image. Moreover, we use the frequency restriction module to guarantees the privacy of users' face images after social network compression. The experimental results show that CSP3Adv achieves better privacy protection for various face recognition models and commercial API interfaces on different datasets.