Broad Learning Inference Based on Fully Homomorphic Encryption

Abstract

Distributed big data computing environments such as machine learning are widely deployed and applied on the cloud. However, since cloud servers can easily access user data, it leads to serious data leakage problems. As a potential technology, Fully Homomorphic Encryption (FHE) is often used in the field of privacy-preserving machine learning. However, in order to reduce the multiplicative depth of FHE, the neural network prunes the number of network layers, resulting in low inference accuracy. A learning model named Broad Learning System (BLS) has the characteristics of shallow model depth and low complexity. Based on this mode, we propose a privacy-preserving inference algorithm with low multiplicative depth, namely broad learning inference based on fully homomorphic encryption. We also extend the algorithm to the BLS model with incremental learning. We implement the privacy-preserving BLS for the first time using TENSEAL’s CKKS scheme, and also verify the effectiveness of BLS inference with incremental learning. Experimental evaluations demonstrate the inference accuracy of 0.928 and 0.672 for datasets of MNIST and NORB, respectively.