FaceIDP: Face-IDentification Privacy under theCompressed Sensing Framework

Abstract

In FaceID era, large number of facial images could be used to breach the FaceID system, which demands effective FaceID privacy protection of the facial images for widespread adoption of FaceID technique. In this paper, to our best knowledge, we take the first step to systematically study such important FaceID privacy issue, under the framework of Compressed Sensing (CS) for fast facial image transmission. Specifically, we develop the Face-IDentification Privacy (FaceIDP) approach to protect the facial images from being used by the adversary to breach some FaceID system. First, a Dictionary Learning neural Network (DLNet) has been developed and trained with facial images database, to learn the common dictionary basis of the facial image database. Then, the encoding coefficients of the facial images are obtained. After that, the sanitizing noise is added to the encoding coefficients, which obfuscates the FaceID feature vector that is used to identify the FaceID. We have also proved that the FaceIDP is $\varepsilon$-differentially private. More importantly, optimal noise scale parameters have been obtained via the Lagrange Multiplier (LM) method to achieve better data utility for a given privacy budget $\varepsilon$. Finally, substantial experiments have been conducted to validate the efficiency of the FaceIDP with two real-life facial image databases, i.e., the LFW (Labeled Faces in the Wild) database and the PubFig database, and the results show that it outperforms other commonly used Differential Privacy (DP) approaches.