Privacy-Preserving Facial Recognition Models: Retaining Learning Without Storing Facial Information

Abstract

At present, most application developers still need to save face images on the server, so they may still be stolen or stolen by service providers. Application service providers can use user facial information in specific visual scenes, but they cannot restrict service providers from using visual information on the server. This paper studies a server-oriented face image privacy protection technology. The research results of this paper are color-based face image perturbation algorithms, which will provide new ideas for solving face detection problems under complex backgrounds and improve the accuracy of detection. The face perturbation algorithm based on feature points can effectively overcome the shortcomings of existing algorithms that only focus on one feature and lack universality. The corresponding mathematical proof is also given in this paper, and the corresponding theoretical basis is given. Through experiments on face databases, we found that the accuracy of using different neural networks to detect images before and after perturbation is between 0.20% and 6.38%. Various different face image quality assessment methods are used and compared with the existing best face perturbation algorithms. The experimental results show that the image quality of the scrambled images is still greatly improved, and compared with the existing algorithms, it has better effects and can meet the needs of data availability.