Abstract Federated Learning(FL) has emerged as a key paradigm for training machine learning models on decentralized data while preserving privacy. However, traditional synchronous, single-server FL architectures suffer from a bottleneck, where system performance is limited by the slowest clients and the central server's capacity. Asynchronous and multiserver approaches have been proposed to improve scalability and efficiency, but they introduce new challenges. In particular, when clients are unevenly distributed across servers, the aggregation process can become heavily skewed by data imbalance, leading to a greater degradation in global model accuracy.

This thesis addresses the problem of client and data imbalance in asynchronous, multiserver FL environment. We propose imbalance-aware aggregation strategies at both the client-server and server-server levels, designed to reduce the dominance of overloaded servers and better reflect heterogeneous client populations.

We implement and evaluate our proposed method on the MNIST dataset under various conditions of data and client imbalance. The MNIST dataset consists of handwritten digit images. The experimental results demonstrate that our approach improves global model accuracy and convergence speed compared to baseline methods, especially under severe imbalance.