Who Trains Matters: Federated Learning under Enrollment and Participation Selection Biases
Federated learning systems assume training participants mirror the target population, but real-world deployments fail this assumption at two critical junctures. Enrollment bias emerges from device constraints and consent rules that exclude entire client cohorts before training begins, while participation bias stems from runtime factors like battery state and network availability that skew which clients actually contribute. This work moves beyond existing round-level participation analysis to examine how both biases compound, reshaping how practitioners must think about model representativeness in distributed settings. For teams deploying FL at scale, the implication is stark: model quality degrades predictably when you ignore who can and will show up.
Modelwire context
ExplainerThe paper's contribution isn't just naming two bias types but showing they compound in ways that make round-level auditing, the current standard practice, structurally insufficient. A model can pass per-round participation checks and still be systematically unrepresentative if the enrolled population was already skewed before training started.
This connects directly to the theme running through the 'SciHorizon-DataEVA' coverage from the same day: the gap between assuming data or participants are fit for training and actually verifying that assumption upstream. DataEVA frames this as a readiness problem for scientific datasets; the federated learning paper frames it as a population validity problem for distributed client pools. Both argue that wasted compute and degraded models trace back to unexamined intake conditions rather than model architecture choices. The electricity forecasting piece also touched this nerve, noting that domain shifts invalidate models trained on populations that no longer reflect current conditions.
Watch whether major FL frameworks like Flower or TensorFlow Federated incorporate enrollment-stage auditing tools within the next 12 months. Adoption there would signal the field treating this as an engineering requirement rather than a research footnote.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsFederated Learning
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.