Achieving Prior-Aligned Utility-Privacy Trade-Off for Data Sharing

Abstract

In privacy-aware data sharing, achieving a prior-aligned trade-off between task utility and sensitive information leakage remains a critical and challenging problem. Conventional approaches typically adopt coarse-grained feature selection strategies, often sharing all attributes or fixed subsets without fine differentiation. Such indiscriminate sharing not only exacerbates privacy leakage risks but also undermines utility due to redundant data perturbation. To address these limitations, this paper proposes a novel, unified utility-privacy modeling framework that facilitates prior, fine-grained, adaptive data sharing under explicit privacy and utility constraints. This framework formulates the trade-off as a constrained optimization problem, where multiple privacy constraints are encoded as weighted penalty terms, enabling flexible prioritization according to scenario-specific requirements. By leveraging an information-theoretic abstraction, both utility gain and privacy loss are estimated via normalized mutual information (NMI), providing a coherent, scale-invariant metric for joint optimization. Unlike traditional methods that indiscriminately share or perturb features, our approach selectively identifies and shares only the most informative data that satisfy privacy constraints, thereby effectively mitigating unnecessary privacy leakage and avoiding utility degradation due to redundant noise addition. The emergent optimisation is achieved through a relaxed dual formulation Lagrangian, yielding an interpretable and computationally efficient mechanism to characterize the utility-privacy trade-off frontier. This work advances the state-of-the-art by enabling precise, task-oriented data sharing policies that dynamically adapt to heterogeneous privacy constraints, paving the way for sophisticated privacy-preserving data exchange and collaborative learning systems.