Unfederated: Open Challenges, Deployment Gaps, and Emerging Directions in Federated Learning

Abstract Federated Learning (FL) promises collaborative and distributed learning, yet remains a largely theoretical construct in practice. Most deployments fail to achieve true federation, as data pipelines quietly re-centralize for quality control, institutions withhold participation without enforceable incentives, and regulatory fragmentation forces siloed rather than collaborative training. These structural failures, not algorithmic issues, explain why FL has not scaled. This review systematically analyzes the full FL lifecycle, from client selection through deployment, exposing how statistical and system heterogeneity, communication overhead, and privacy-utility trade-offs interact to compound failure at scale. We identify open challenges and systemic barriers at each lifecycle stage, surveying state-of-the-art responses including personalization via meta-learning, secure aggregation, communication sparsification, and Byzantine-robust optimization, while identifying where each falls short. Emerging directions, including reinforcement learning for adaptive coordination, graph-based trust propagation, parameter-efficient federated foundation models, and standardized benchmarks, are examined as principled paths forward. Closing the gap between FL’s promise and practice demands enforceable incentive mechanisms, cross-institutional trust frameworks, and evaluation standards designed around operational reality rather than lab settings.