Face Privacy Protection Based on Attribute Manipulation

Abstract

Recent studies have shown the possibility of inferring soft biometric attributes from individual facial images, such as age, gender, and race. This has aroused people's concern about privacy. To solve this problem, we designed a generation confrontation network to achieve the purpose of privacy protection by confusing or preserving the soft biometrics of face images and preserving the identity matching function of the original face. Our method can complete the conversion between multiple facial attributes by using only one generation model. Users can flexibly select the retention or modification of facial attributes without affecting the accuracy of identity matching. A large number of experiments show the effectiveness of our method.