Shielding Collaborative Learning: Mitigating Poisoning Attacks through Client-Side Detection

Lingchen Zhao (Wuhan University), Shengshan Hu (Huazhong University of Science and Technology), Qian Wang (Wuhan University), Jianlin Jiang (Wuhan University), Chao Shen (Xi'an Jiaotong University), Xiangyang Luo (State Key Laboratory of Mathematical Engineering and Advanced Computing), Pengfei Hu (VMware, Inc.)

Published Oct 29, 2019 · arXiv:1910.13111

cs.CRcs.LG

Abstract

This paper presents a novel defense scheme that delegates the detection of anomalous updates to clients in collaborative learning to mitigate poisoning attacks in both IID and non-IID settings while protecting client-level privacy by integrating differential privacy.

arXiv abstract PDF