RECAST Reconstructs Black-Box Models with Limited Data and Access.

Understanding and auditing black-box machine learning models is crucial for ensuring fairness and accountability, especially in sensitive decision-making systems. Counterfactual explanations (CFs) offer a way to interpret models by showing minimal input changes that alter outcomes. However, using CFs for model reconstruction often faces challenges such as decision boundary shifts, overfitting, and the need for continuous online access to the target model. Researchers have introduced RECAST (REconstruction via Counterfactual-Aware waSserstein opTimization), a behavioral surrogate model designed to overcome these limitations. RECAST leverages Wasserstein barycentric prototypes and incorporates CFs as valuable, albeit less representative, samples for both classes. This approach helps maintain high surrogate fidelity even when data is scarce and online query access is restricted during the reconstruction phase. A significant benefit of RECAST is its ability to facilitate systematic group fairness diagnostics, making it a powerful tool for auditing opaque systems. Experimental results on real-world datasets confirm that RECAST achieves high fidelity and query efficiency, delivering stable results even under noisy and limited access conditions, thereby improving the reliability of model auditing.