"Persistence Landscapes Across Privacy Budgets for Explanation Methods Across Differential Privacy Mechanisms"

"This dataset\/package supports experiments studying how post-hoc explanation methods change under differential privacy in a tabular credit-risk setting using the HELOC benchmark. It contains the processed experimental artifacts needed to reproduce stability analyses across privacy budgets for multiple differential privacy mechanisms (e.g., additive-noise perturbations, randomized response, DP-SGD, differentially private synthetic data generation, and DP-PCA-style transformations) and multiple explanation methods (e.g., LIME, KernelSHAP, Occlusion, Vanilla Gradient, Gradient\u00d7Input, Integrated Gradients, and SmoothGrad). For each (privacy mechanism, explainer, privacy budget) condition, the package provides outputs for a Mapper-based topological data analysis pipeline that converts per-instance attribution vectors into persistence diagrams and persistence landscapes, enabling quantitative comparison of explanation \u201cstructure\u201d as privacy noise increases. The release is intended to facilitate reproducible research on privacy\u2013explainability interactions and to provide benchmark artifacts for future work on stability metrics, robust explanation auditing, and privacy-preserving transparency.\u200b"