Optimizations for federated learning systems
Xinyu Zhou · 2025
With growing concerns over data privacy and increasing regulatory constraints, centralized machine learning systems face significant limitations in handling sensitive user data. Federated Learning (FL) offers a promising alternative by enabling collaborative model training across distributed devices without sharing raw data. However, FL also introduces several challenges—especially in wireless and mobile environments—such as limited communication and computation resources, data and device heterogeneity, potential privacy leakage through model updates, and the need for effective user incentives. This thesis focuses primarily on addressing the system-level challenges related to communication and computation limitations, which severely impact the deployment and efficiency of FL systems. These include bandwidth bottlenecks from model updates and energy costs associated with local training on resource-constrained devices. Moreover, to address these challenges, this thesis mainly focuses on resource allocation optimization for FL in combination with different application scenarios, specifically mobile augmented reality (MAR) and semantic communication (SemCom). First, an FL-assisted MAR system is investigated. To jointly optimize energy consumption, latency, and model accuracy, we formulate and solve a weighted resource allocation problem. A foundational resource allocation algorithm is developed to allocate bandwidth, transmission power, CPU frequency, and video resolution for each user device. Second, this thesis extends the above framework by introducing a new channel access scheme and integrating user experience metrics to further enhance system optimization. A corresponding resource allocation strategy is proposed, and empirical results demonstrate improved performance over conventional baselines under diverse configurations. Finally, a more advanced FL-assisted SemCom scenario is explored. The practical performance feedback is used to refine the FL training and jointly optimize the overall resource allocation strategy. In conclusion, this thesis systematically addresses the key bottlenecks in FL—limited bandwidth, energy constraints, and resource heterogeneity—through novel system models and optimization algorithms tailored for future intelligent communication systems. Theoretical analysis and simulation results confirm the effectiveness of the proposed methods across various scenarios.