Fully homomorphic faces

Face recognition is a prominent application of image processing. It is also a very sensitive application, and privacy concerns have been lately raised and tackled in several recent papers dealing with privacy-preserving face recognition systems. Nevertheless, the presented systems either use the knowledge of some information derived from the database templates in order to perform the recognition or require several interaction rounds between client and server. In this paper, we propose a private system that can cope with a simple verification algorithm executed in the server without interaction (using a quasi-fully homomorphic encryption and an efficient face features representation with Lloyd-Max quantized Gabor jets), in which both the templates and the queried face are encrypted; we show its performance in terms of time complexity and size of transferred encryptions, as well as in verification accuracy with respect to the non-private system. This opens the door to completely private and noninteractive outsourcing of face recognition.