Contribution of Non-scrambled Chroma Information in Privacy-Protected Face Images to Privacy Leakage

Abstract

To mitigate privacy concerns, scrambling can be used to conceal face regions present in surveillance video content. Given that lightweight scrambling tools may not protect chroma information in order to limit bit rate overhead in heterogeneous usage environments, this paper investigates how the presence of non-scrambled chroma information in face regions influences the effectiveness of automatic and human face recognition (FR). To that end, we apply three automatic FR techniques to face images that have been privacy-protected by means of a layered scrambling technique developed for Motion JPEG XR, testing the effectiveness of automatic FR and layered scrambling using various experimental conditions. In addition, we investigate whether agreement exists between the judgments of 32 human observers and the output of automatic FR. Our experimental results demonstrate that human observers are not able to successfully recognize face images when simultaneously visualizing scrambled luma and non-scrambled chroma information. However, when an adversary has access to the coded bit stream structure, the presence of non-scrambled chroma information may significantly contribute to privacy leakage. By additionally applying layered scrambling to chroma information, our experimental results show that the amount of privacy leakage can be substantially decreased at the cost of an increase in bit rate overhead, and with the increase in bit rate overhead dependent on the number of scalability layers used.