Fed-ARDP: An Adaptive Reputation-Driven Federated Learning Framework Using Renyi Differential Privacy

Abstract

Federated Learning (FL) approach can promote collaborations of training the predictive model among distributed participants without sharing raw data, to maximize the privacy protection in the heterogeneous networked environment. Conventional FL architectures often suffer from issues like indiscriminate addition of noise in the privacy terms, uneven client participation and unnecessary multiplication of communication rounds there by potentially compromising model fidelity and training through. In response, this study introduces Fed-ARDP, a framework called Adaptive Renyi Differential Privacy in which privacy perturbation is modulated dynamically according to the contributory weight of each client. The methodology quantifies client contributions using a hybrid contribution scoring mechanism combining validation accuracy and cosine similarity between local and global model updates and dynamically adjusts Rényi Differential Privacy (RDP) noise according to the resulting reputation scores. Furthermore, a reputation-weighted aggregation scheme is employed to improve the reliability of global model updates. The proposed framework is evaluated against FedAvg, FedProx, Krum, Bulyan, and Fixed-DP under IID, Non-IID, and adversarial settings to show that Fed-ARDP achieves better model accuracy, has stable convergence, and lower training latency as compared to the baseline federated learning strategies. Therefore, the proposed framework achieves a balanced trade-off between privacy, efficiency, and robustness in multi-party federated learning environments.