Zero-Trust Empowered Decentralized Security Defense against Poisoning Attacks in SL-IoT: Joint Distance-Accuracy Detection Approach

Swarm learning (SL) exploits the blockchain to realize a federated and decentralized learning, which is very suitable for internet of things (IoT). Different from FL using central server to update global parameter, SL using edge node (header) to do that. However, poisoning attack is also an unresolved problem to SL. Because if header is malicious, it can pollute global parameter more easily than edge nodes. Moreover, there are following important limitations in existing defense schemes for FL, which cannot be used in SL directly. First, existing defense schemes focus on building a whitelist, which obstructs the decentralization because it can just provide decentralization in honest nodes instead of all of nodes. Second, existing schemes just consider poisoning attacks from edge nodes, they cannot defend attacks from header. Third, most existing schemes will let server execute the defense algorithm, but in SL, malicious header can return wrong defense results to deceive managers. To address above challenges, in this paper, we propose a protection system that leverages the concept of zero-trust architecture for SL, which achieves continuous risk calculation, analysis of learning behavior and abnormal parameter detection based on Manhattan distance and accuracy difference of parameters. We also evaluate the performance in the presence of random and customized malicious edge nodes. Experimental results demonstrate that our scheme can achieve higher accuracy than the other existing schemes.