Fed-DCSRW: a privacy-preserving, dynamic client selection framework for heterogeneous federated learning via roulette wheel mechanism

Abstract We introduce Fed-DCSRW, a federated-learning framework that combines three pillars: (i) a decentralized, data-parallel client-clustering stage that computes centroids in parallel to scale and compute efficiently across heterogeneous clients; (ii) a centroid-based, noise-tolerant Roulette-Wheel client-selection strategy; and (iii) end-to-end differential privacy on both loss reports and gradient updates. The parallel clustering phase partitions the distance computations across multiple nodes, rapidly deriving similarity-based centroids without centralized bottlenecks. These centroids then anchor a loss-aware Roulette Wheel that adjusts each round’s participation probabilities: clients with lower (possibly noisy) losses retain a non-zero chance of being chosen, while the overall selection proportion is dynamically adjusted based on real-time global loss trends, enabling the system to respond to performance changes and prevent training stagnation. Experimental results on non-IID datasets show that Fed-DCSRW achieves faster empirical convergence and higher test accuracy compared to standard methods such as FedAvg, Standard Roulette Wheel Selection (RWS), and Utility-Based Probabilistic Selection (Oort). These results demonstrate the effectiveness of a privacy-preserving, centroid-driven selection strategy in addressing data heterogeneity while maintaining strong model performance.