Semantic utility-driven client selection for task-oriented split federated learning

Task-oriented split federated learning (TOSFL) is a collaborative intelligence framework that integrates split learning and semantic communication, which holds great promise for edge intelligence. However, its practical application is hampered by inherent information asymmetry, which severely affects training efficiency and security. In this paper, we introduce the semantic utility-driven client selection (SUCS) framework. At the core of SUCS is the semantic utility index (SUI), a novel metric designed to quantify the contribution of clients in two orthogonal dimensions: (1) instantaneous data value, which measures how client data reduces global model uncertainty from an information-theoretic perspective; and (2) long-term model reputation, which evaluates reliability by tracking historical performance. Based on SUI, we develop a dynamic selection and feedback mechanism to optimize client participation and guide local data sampling. Comprehensive experiments on multiple datasets and two tasks show that SUCS improves test accuracy by 5.5%, reduces the number of communication rounds to converge by 28%, and exhibits significant robustness to data redundancy and adversarial attacks compared to a random selection baseline.