ProFL-DARL: Privacy-Preserving Robustness Federated Learning with Dynamic Aggregation Based on Reinforcement Learning

When confronted with highly non-independent and identically distributed (Non-IID) data, inconsistent client quality, and significant privacy encryption and communication overhead, federated learning systems often suffer from severe global model bias, suboptimal local personalized performance, and high privacy leakage risks. To address this challenge, this paper proposes a collaborative computing framework that separates classification servers and aggregation servers. It designs reinforcement learning agents to achieve automatic gradient similarity identification, online grouping, and weight allocation while filtering out anomalous clients. The integration of RLWE homomorphic encryption with gradient masking technology substantially enhances security against adversarial collusion attacks. Comparative experiments on Non-IID datasets demonstrate that this approach significantly improves local model accuracy for clients with limited data, while keeping communication, encryption, and differential privacy overhead within system-acceptable bounds.