BackDFL Benchmark Exposes Decentralized Federated Learning Vulnerabilities

Decentralized Federated Learning (DFL) is often lauded for its trust-free collaborative learning model, replacing central servers with peer-to-peer model exchanges. However, this architectural shift introduces new security challenges, particularly concerning backdoor attacks where malicious participants can implant hidden behaviors while maintaining apparent normal performance. A new benchmark, BackDFL, argues that the robustness of DFL against such attacks has been significantly overestimated. Existing studies, according to the researchers, have relied on simplified threat models, non-adaptive adversaries, and inconsistent evaluation protocols, leading to an incomplete understanding of DFL's security posture. BackDFL addresses these limitations by providing a unified framework for systematically evaluating DFL under realistic and adaptive backdoor attacks. Through extensive experiments, BackDFL uncovered critical failure modes in decentralized learning. The results show that state-of-the-art Byzantine-robust DFL methods and even adapted federated learning backdoor defenses fail when malicious participation rates are as low as 15%, especially in heterogeneous settings. Furthermore, robustness varied substantially across different communication graph topologies. This benchmark highlights that DFL is more susceptible to backdoor attacks than previously thought, necessitating more robust defense mechanisms.