Advanced Backdoor Attacks on Decentralized Federated Learning

Decentralized Federated Learning (DFL) enhances robustness and privacy by removing the central server, but this decentralized setting also introduces new vulnerabilities. While backdoor attacks in centralized federated learning are relatively well-studied, their extension to DFL is still in its infancy. The lack of a trusted coordinator, the peer-to-peer aggregation, and the dynamic topologies in DFL open new possibilities for stealthy, adaptive, and topology-aware backdoor attacks.

This thesis will investigate advanced backdoor attack strategies specifically tailored for DFL environments. The student will design novel attack vectors that exploit structural properties of decentralized topologies (e.g., ring, mesh, random, or real-world graphs) and aggregation mechanisms. Possible research directions include adaptive trigger design, poisoning under constrained communication, or topology-aware collusion among malicious nodes.

None