Your Neighbors Know: Leveraging Local Neighborhoods for Backdoor Detection in Decentralized Learning

A new framework named Argus has been introduced for backdoor detection in decentralized learning (DL), a machine learning paradigm where nodes collaborate without a central server. Argus allows honest nodes to analyze model updates locally and share potential backdoor triggers with their neighbors, using a structural similarity metric to distinguish between true backdoors and false alarms caused by data heterogeneity.

This development is significant as it addresses the vulnerabilities of decentralized learning to backdoor attacks, enhancing the security and reliability of collaborative machine learning systems without requiring a central coordinator or prior knowledge of triggers.