FedDCDS: a dual-phase secure federated learning framework with dynamic clustering and adaptive delivery

Abstract

FedDCDS introduces a novel dual-phase secure federated learning framework designed to address the intertwined challenges of statistical heterogeneity, system heterogeneity, and security threats in large-scale edge computing environments. By integrating a dynamic spectral clustering mechanism, the framework dynamically groups clients with similar computational capabilities, network conditions, and data distributions into clusters, mitigating Non-IID data drift and enhancing intra-cluster aggregation efficacy. The dual-phase secure delivery protocol employs threshold secret sharing during client uploads to protect gradient confidentiality and leverages adaptive model slicing with covert transmission during server-to-client downloads, ensuring end-to-end security while minimizing communication overhead. An adaptive noise injection mechanism balances privacy and utility by injecting exponentially decaying Gaussian noise into global updates, satisfying differential privacy constraints without compromising convergence accuracy. Extensive experiments on MNIST and CIFAR-10 datasets demonstrate FedDCDS’s superiority: it achieves global accuracy improvements of 76.93% and 96.84%, reduces cross-group aggregation latency by 40–49%, and suppresses model inversion attacks to 6.9–12.8% reconstruction accuracy, significantly outperforming state-of-the-art baselines in scalability, robustness, and privacy preservation.