New Backdoor Attack Threatens Decentralized Federated Learning

Semantic backdoor attacks in federated learning (FL) are challenging to detect due to their subtlety, but their effectiveness can be diminished by sample-dependent placement and repeated model aggregation. This challenge is amplified in decentralized federated learning (DFL), where local models are frequently mixed through topology-dependent peer aggregation. This paper presents CACTUS, a new mask-guided semantic clean-label backdoor attack specifically designed to overcome these hurdles in DFL environments. CACTUS operates by transforming label-consistent semantic pairs into targeted representation shifts. It employs mask-guided, modality-specific operators to isolate trigger effects, link them across different samples, and then apply these shifts counterfactually to clean non-target embeddings before peer aggregation occurs. Experiments across speech, text, tabular, and image tasks, under nine different aggregation rules, demonstrated CACTUS's efficacy. With only 30% malicious nodes, it achieved a mean attack success rate (ASR) of 51.2% on Speech Commands and the highest mean ASR among tested attacks in three out of four modalities, proving its ability to propagate backdoors through repeated DFL aggregation.