ApaPRFL: Robust Privacy-Preserving Federated Learning Scheme Against Poisoning Adversaries for Intelligent Devices Using Edge Computing
Shaojun Zuo, Yong Xie, Libing Wu, Jing Wu
Abstract
The large amount of data collected by intelligent devices in consumer electronics cannot be fully utilized because it involves a lot of privacy information. At present, researchers propose many security protection schemes, among which the scheme using edge computing architecture attracts much attention. However, existing schemes cannot simultaneously address security, efficiency, and robustness, especially in the case of intelligent devices dropout. Therefore, we propose an intelligent device data secure federated learning scheme using edge computing architecture named ApaPRFL. ApaPRFL is based on the gradient strong privacy-preserving method using secure secret sharing. It leverages the property of high regional similarity to ensure system stability even when the end devices (intelligent devices) dropout. Additionally, it improves the efficiency of poisoning detection and reduces error rates. The performance of ApaPRFL is evaluated on two real datasets. Experimental results demonstrate that ApaPRFL is more effective in countering two typical poisoning attacks compared to similar schemes.