FedStream: A Federated Learning Framework on Heterogeneous Streaming Data for Next-Generation Traffic Analysis
Naiyu Wang, Xuan Li, Zhitao Guan, Shuai Yuan
Abstract
Network traffic analysis is a key paradigm for the development of next-generation networks. Extensive deployment of network devices enables the collection of real-time traffic data for statistical analysis, thus providing timely feedback on unexpected situations that may arise. However, traditional data-driven methods fail to capture the unique properties of such data, which result in models with poor performance, low efficiency, and potential privacy risks. Though much of the pertinent work has focused on the federated learning as the alternative, the heterogeneity of both data distributions and arrival patterns have not been taken into account. We therefore define the dual heterogeneity challenge and conduct extensive experiments to explore the impact of the challenge on existing optimization algorithms. Subsequently, to address the challenge, we propose H_strSAGA in local optimization stage and develop a federated learning framework FedStream, where the various computing capabilities and communication bandwidths are considered by sequencing devices during the aggregation process. Furthermore, as the heterogeneity among devices increases, synchronous training may face potential challenge with stragglers, we additionally incorporate an asynchronous aggregation algorithm. Experimental studies using real-world datasets have demonstrated significant improvement on the convergence of the global model and the performance of model generalization.