Multi-Party Privacy-Preserving Faster R-CNN Framework for Object Detection

Abstract

Faster region-based conventional neural network (Faster R-CNN) is a common algorithm for object detection that identifies the object and their location information through three steps: feature extraction, region proposal network and classification. However, there are data privacy issues in the training and prediction of Faster R-CNN. Thus we design a multi-party privacy-preserving Faster R-CNN framework for object detection named SPFR. Specifically, we extend the existing sub-protocols and achieve high-precision division, exponentiation and logarithm calculation through the idea of blinding and oblivious transfer (OT) protocols. Then we design a series of key privacy-preserving protocols that satisfy the secure computational requirements based on the above-mentioned sub-protocols and prove that these protocols are security in the semi-honest adversary model. Finally the proposed protocols are implemented in Pytorch, and the experimental findings demonstrate the efficiency of the protocols in Faster R-CNN.