Backdoor Poisoning Mitigation for Federated LoRA via Structural Adaptive Control
Abstract
Large pre-trained models are increasingly adapted on user devices, but direct fine-tuning in federated settings is constrained by computation, memory, and communication limits. These constraints have motivated the use of parameter-efficient fine-tuning methods, among which LoRA is widely adopted by concentrating adaptation in low-rank parameters. While this design improves efficiency, it also changes the security setting. In federated LoRA, backdoor behavior can exploit the low-rank adaptation structure and persist across communication rounds, making defenses designed for full-model updates less reliable. To address this gap, we propose StrAP, a server-side defense tailored to federated LoRA fine-tuning. StrAP combines reliability control before aggregation with structure-aware regulation after aggregation. Before aggregation, client updates are softly weighted based on their alignment with a clean validation gradient and tracked across rounds using a stateful trust mechanism, which suppresses persistent anomalies while allowing benign recovery. After aggregation, StrAP analyzes the LoRA parameter structure to identify suspicious rank channels using robust cross-client statistics and applies utility-calibrated structural actions guided by clean validation loss. Extensive experiments show that StrAP consistently reduces backdoor attack success rates with minimal impact on clean accuracy. Our code is available at https://github.com/HwhCN-ui/StrAP.