Context-Aware Client Collaboration in Federated Learning with non-IID data
PhD Dissertation Proposal by: Soroush Ziaeinejad
Date: Thursday, November 27, 2025
Federated Learning (FL) suffers significantly in real-world deployments due to the high degree of statistical heterogeneity across clients. While many existing strategies rely on random sampling or simplistic heuristics such as data size or recent accuracy, these approaches often fail to capture the evolving utility of each client throughout training, especially in non-IID settings. This proposal introduces a curriculum-driven strategy for client selection that adaptively orders client participation based on a combination of performance-aware, data-aware, and optimization-aware criteria. By incorporating metrics such as local loss and label-distribution entropy, the method adapts the participation schedule as training progresses. This creates a smoother optimization path and more balanced contributions from heterogeneous clients. We formalize a multi-criteria strategy for curriculum-based selection and evaluate it using MedMNIST benchmarks under various non-IID scenarios. We compare the approach against common baselines, including random selection and accuracy-based heuristics, using metrics such as accuracy, fairness, training stability, and participation frequency. The expected outcome is a principled and efficient selection mechanism that improves both performance and fairness in federated systems operating under significant statistical heterogeneity