Efficient Adaptive Asynchronous Federated Learning Based on Differential Privacy

Asynchronous Federated Learning (AFL) is an efficient and secure distributed machine learning paradigm designed to tackle challenges such as data silos and client heterogeneity. However, practical implementations often suffer from sporadic client participation due to suboptimal network conditions or varying device quality, resulting in significant latency for model updates. Additionally, data privacy is also a security issue that cannot be overlooked in Federated Learning. Nevertheless, existing adaptive federated optimization algorithms fall short in effectively addressing the dual challenges of high client-side latency and robust privacy protection. To address these challenges, we propose DPAAFL, an adaptive asynchronous federated learning framework with per-client differential privacy. DPAAFL improves the accuracy–efficiency trade-off by adopting a dual-event triggering rule that admits only sufficiently informative updates with acceptable latency, and by mitigating the impact of staleness through delay-adaptive aggregation; meanwhile, a buffered asynchronous update scheme reduces the communication overhead by aggregating updates in batches. In addition, each client perturbs its clipped update with calibrated Gaussian noise before transmission, providing differential privacy protection against privacy inference attacks under the honest-but-curious threat model. We provide theoretical analyses for both convergence and privacy. Extensive experiments demonstrate that the proposed scheme not only reduces the communication overhead and enhances system robustness, but also significantly improves the training efficiency and ensures the accuracy of the model. In comparison with contemporary state-of-the-art algorithms, the proposed scheme improves model accuracy and training efficiency by up to 3.5% and 29.3%, respectively.