DuoShapley: Adaptive and Scalable Shapley Value Approximation for Federated Learning

TMLR Paper7264 Authors

Abstract: Federated Learning (FL) enables collaborative model training across decentralized users while preserving data privacy, but it also raises a fundamental challenge: how to efficiently and reliably quantify individual user contributions to the global model. The Shapley value (SV) provides a principled game-theoretic framework for contribution valuation, yet its exact computation is prohibitively expensive in realistic FL systems. Existing SV approximation methods face a trade-off between scalability and estimation fidelity, particularly under heterogeneous data distributions. In this work, we propose DuoShapley, an efficient and adaptive SV approximation tailored to large-scale FL that adaptively balances two complementary orders: Solo, capturing individual contributions, and Leave-One-Out (LOO), capturing marginal contributions relative to the full coalition. By adaptively weighting them during training based on the alignment between local and global model updates, DuoShapley achieves both computational efficiency and accurate contribution valuation across diverse FL scenarios, from independent and identically distributed (IID) to non-IID. Beyond contribution measurement, DuoShapley enables downstream applications such as robust user selection in the presence of users with noisy data, by prioritizing users with high estimated contributions. Such selective participation leads to enhanced robustness to noisy and low-quality updates, and reduced communication overhead. Extensive experiments show that DuoShapley is both computationally efficient and effective across diverse data distributions. Hence, DuoShapley provides a practical and scalable solution for evaluating and leveraging user contributions in FL.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Ju_Sun1

Submission Number: 7264