New research quantifies how fragile federated learning's accuracy promise gets when real-world device data stops behaving itself.
Researchers designed a battery of experiments comparing centralized, local, and federated training approaches for Human Activity Recognition - the task of classifying whether a person is walking, running, or sitting based on sensor data. They applied FedAvg, the standard algorithm for aggregating model updates across distributed devices, and measured how well it balanced two competing goals: fitting each user's data closely (personalization) versus working well across all users (generalization). Under normal conditions, FedAvg edged out centralized training on both counts. When the researchers simulated shifting class distributions - where some devices mostly see one type of activity while others see a different mix - the advantage became considerably less clear.
The result matters because federated learning's appeal in health and fitness applications rests on a straightforward premise: train on the phone, share only model weights, preserve privacy. But user behavior is inherently uneven. A fitness tracker worn by a marathon runner captures a very different slice of activity than one worn by a desk worker, and that mismatch is exactly the condition where FedAvg's performance guarantees start to soften.
The uneven-data problem in federated learning - sometimes called non-IID, where IID stands for "independent and identically distributed" - is well-documented in the field. This paper does not solve it, but it gives engineers a cleaner picture of how much it hurts in a concrete, practical domain.