The Critical Role of Client Selection in Cross-Silo Federated Learning: Challenges, Approaches, and Opportunities

Cross-silo federated learning is a promising paradigm that enables organizations to jointly train machine learning models while preserving the privacy of sensitive data within institutional boundaries. Cross-silo settings involve fewer strategic institutional actors with significant computational resources and data repositories than cross-device federated learning, which links many devices with limited resources. In this regard, the process of selecting which organizations participate in each training round, known as client selection, is a crucial yet understudied decision point that significantly impacts model performance, system efficiency, privacy assurances, and the sustainability of long-term cooperation. By conducting a thorough comparative analysis, the important research gaps were identified, such as the lack of empirical validation in actual cross-silo deployments, the lack of attention to multi-objective optimization, the limited incorporation of regulatory constraints, and the inadequate modeling of strategic organizational behaviour. Through a methodical analysis, this review paper examines the state of client selection techniques in cross-silo federated learning. In order to distinguish cross-silo selection from its cross-device counterpart, a new taxonomy is created that divides selection techniques into resource-based, data-centric, and strategic frameworks. Although the majority of the literature to date has concentrated on technical efficiency in cross-device settings, this review shows that the organizational nature of cross-silo participants calls for more complex strategies that strike a balance between incentive compatibility, coalition stability, regulatory compliance, and immediate model performance.