HeteRo-Select

HeteRo-Select is a client selection and aggregation framework for federated learning under extreme statistical and system heterogeneity. It addresses the challenge of training a single global model across participants with vastly different data distributions and compute capacities.

Key Contributions

• Variance-aware client scheduling that prioritizes under-represented data partitions
• Heterogeneity-aware aggregation that downweights stale or outlier updates
• Rigorous evaluation on CIFAR-10/100 with non-IID splits and heterogeneous hardware profiles

Links

• arXiv preprint
• GitHub