Improving the accuracy and stability of privacy-aware Shapley data valuation

Abstract The integration of privacy-preserving mechanisms like differential privacy into data valuation based on Shapley values introduces systematic noise, significantly degrading the accuracy and stability of estimates—a challenge not adequately addressed by existing methods. To address this, we propose the Privacy-Aware Shapley Valuation (PASV) framework, which combines a bias calibration module based on noise propagation modeling, designed to correct systematic deviations caused by noise propagation, with a variance-aware dynamic sample reuse scheduler that adaptively selects high-confidence historical samples for reuse for enhanced stability. Comprehensive experiments on multiple benchmark datasets under varying privacy budgets and federated learning settings (including both IID and non-IID data partitions) were conducted to validate our approach. Results demonstrate that PASV reduces estimation bias by an average of 18.7% and improves ranking stability by over 35% compared to standard private Shapley methods, while maintaining identical (ε, δ)-differential privacy guarantees. The bias calibration is realized via a lightweight online correction factor fitted on synthetic non-sensitive datasets to avoid privacy leakage, and the dynamic sample reuse scheduler minimizes the upper bound of Shapley estimation variance under privacy budget constraints.This work provides a practical solution for trustworthy data valuation under privacy constraints and advances the theoretical understanding of noise management in cooperative game-theoretic frameworks, and also lays a foundation for data pricing and incentive mechanism design in real-world data markets and federated learning systems.