Fed-DCR:基于双域协同检测的联邦学习后门防御方案OA

Federated learning has been shown to be vulnerable to backdoor attacks,while existing defense methods mainly rely on single statistical features in the parameter space.Under non-IID conditions,the overlap between benign and mali-cious update features often leads to degraded detection performance,while dedicated defense schemes for low-rank fine-tuning scenarios remain underexplored.To address these issues,Fed-DCR,a backdoor defense scheme for federated learn-ing based on dual-domain synergistic detection,was proposed.In the frequency domain,the discrete cosine transform(DCT)was used to decouple client updates,while semantic alignment and spectral density clustering were introduced to improve the separability between benign and malicious updates.In the parameter domain,low-rank adaptation(LoRA)spectral consistency detection was employed to identify anomalous deviations in the low-rank subspace,while causal prob-ing was integrated to actively expose hidden backdoors.Experimental results show that Fed-DCR effectively reduces the success rates of multiple backdoor attacks on both the CIFAR-10 vision task and the Llama-2-7B-based federated LoRA fine-tuning task.Fed-DCR provides a favorable balance between backdoor mitigation and main-task performance across most experimental settings.