Revealing the influence of participant failures on model quality in cross-silo Federated Learning

A team of researchers including Fabian Stricker, David Bermbach, and Christian Zirpins has published a comprehensive study examining a critical gap in Federated Learning (FL) reliability. Their paper, "Revealing the influence of participant failures on model quality in cross-silo Federated Learning," systematically investigates how crash failures, network partitioning, and participant dropouts impact ML model outcomes in distributed training environments. The research addresses a fundamental challenge: while FL promises privacy-preserving collaborative training by keeping data local, its distributed nature makes it inherently susceptible to failures that could compromise model validity, stability, and reproducibility.

Through extensive experiments across diverse data types—including image, tabular, and time-series datasets—the researchers analyzed how participant absence affects model performance. They examined multiple influencing factors such as data skewness, different availability patterns, and various model architectures. One of the most significant findings reveals that data skewness has a particularly strong impact, often leading to overly optimistic model evaluations and, in some cases, even altering the effects of other influencing factors. The study also explores scenario-specific aspects like the utility of the global model for missing participants, providing detailed insights that could inform more robust FL system design and deployment strategies.