Privacy-Preserving Distributed Face Recognition System Based On Garbled Circuit And Deep Learning

In the Distributed Face Recognition System(DFRS),the client acquires the face data and sends it to the server,which uses the face template stored in the database to match and feeds the result back to the client.Due to the need for information transfer,in addition to malicious third-party threats,DFRS also faces the problem of mutual distrust among participating parties.Without a trusted third party's involvement,a more effective approach to these problems is to introduce a secure two-party computing protocol(STPC)in cryptography.The STPC protocol has the following characteristics:(1)The inputs of the parties involved in the calculation are confidential;(2)The participants can get the correct output;(3)The participants cannot obtain additional valid information except their own input and output.However,there are compatibility problems between STPC protocol and face recognition algorithm:(1)Due to the limitation of STPC protocol,the choice of face recognition algorithm is limited,resulting in low accuracy;(2)The faces recognition algorithm with higher accuracy generally has a high degree of nonlinearity,and it is difficult to combine with the STPC protocol.So exploring how to introduce compatible STPC protocol in DFRS system has certain research value.In order to ensure the accuracy of the DFRS system,the face recognition algorithm based on deep learning(DL)is adopted.The DL-based DFRS system is divided into an offline training phase and an online identification phase.This article focuses on the privacy-protection issues of the DFRS system in the online identification phase and conducts the following research:(1)The compatibility between STPC and DL is solved by selecting the appropriate STPC protocol and DL nonlinear activation function;(2)Design and build privacy-protected DFRS based on GC and DL,achieving 95% accuracy,taking into account security and accuracy;(3)Aiming at the efficiency problem of Privacy-Preserving DFRS,an improved scheme is proposed: the DL model is split,then a part of the calculation is completed in the client's preprocessing stage,and the rest of the calculations execute the GC protocol.The time spent on the entire system was reduced from 3 minutes in the original plan to 5.7 seconds.