On Improving Federated Learning Training Efficiency Over Mobile AI Systems
Yixuan Li, Xiaoqi Qin, Miao Pan, Jiacheng Wang, Dusit Tao Niyato, Ping Zhang · IEEE Communications Magazine · 2025
The rapid advancements in machine learning, wireless communication, and device hardware have created unprecedented opportunities for artificial intelligence (AI) applications on mobile AI systems. Federated learning (FL), a privacy-preserving distributed AI framework, is crucial for the final stage of AI deployment. However, implementing FL on mobile AI systems faces significant challenges, primarily due to device heterogeneity and hardware compatibility issues. Ignoring these real-world constraints hinders the performance of advanced FL algorithms in practical deployments. In this article, we explore methods for modeling FL performance metrics, including training delay, energy consumption, and training efficiency, over mobile AI systems while highlighting key differences from traditional simulation approaches. We then examine the challenges of device heterogeneity and hardware compatibility limitations in FL implementation. Furthermore, we present efficient FL design over mobile AI systems, demonstrating their effectiveness through real-world deployments on devices. Finally, we discuss potential research directions for further enhancing FL efficiency in mobile AI systems from a system-level implementation perspective.