Serverless Federated Learning in Multi-Cell Scenarios Robust to Test-Time Distribution Shifts

This paper introduces a novel serverless Federated Learning (FL) framework designed for multi-cell environments, addressing the significant limitations of traditional centralized FL methods, especially in dynamic settings with frequent test-time distribution shifts. Unlike conventional FL approaches that rely on a central cloud server, our framework decentralizes model training and aggregation entirely to edge servers, thereby reducing communication latency, improving scalability, and enhancing data privacy. The proposed framework incorporates additional hyperparameters of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\alpha$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\beta$</tex> to control the tradeoff between personalization and generalization for the model aggregation process, while ensuring robust performances in out-of-distribution (OOD) tasks. Furthermore, the framework exhibits adaptability across varying numbers of clients in overlapping regions, making it a scalable and effective solution for real-world FL applications in environments with diverse and unreliable connectivity. Experimental results on the CIFAR-10 dataset demonstrate that our framework outperforms existing serverless and centralized FL methods, achieving superior personalization and accuracy even under challenging conditions.