Secure and Robust Aggregation for Federated Learning with Dynamic User Update

Abstract

Existing secure aggregation schemes in federated learning (FL) face challenges related to detecting poisoning attacks and managing dynamic membership updates. To address these limitations, this article proposes a robust and dynamic aggregation framework for FL (RDFL), a robust and dynamic aggregation framework for FL. RDFL integrates a trimmed median algorithm with noninteractive range zero-knowledge proofs, providing a tunable mechanism for detecting abnormal updates. Client behavior is evaluated through a dynamic reputation scoring module, with malicious clients being added to a revocation list. By incorporating revocable attribute-based encryption, RDFL supports dynamic user management, ensuring that only authorized participants can access or update the global model. In addition, RDFL employs Shamir’s secret sharing and a pseudorandom double-masking scheme to maintain aggregation accuracy and protect communication privacy despite client dropouts. Experimental evaluations on the Extended MNIST (EMNIST) and CIFAR-100 datasets demonstrate that RDFL achieves strong security, communication efficiency, and model accuracy, making it suitable for FL involving a large number of clients with dynamic participation.