Reproducibility Package for "Auditing Membership Leakage in Educational Early-Warning Systems: Cross-Dataset Evidence on Data Minimization, Output Restriction, and Differential Privacy"

This reproducibility package accompanies the manuscript “Auditing Membership Leakage in Educational Early-Warning Systems: Cross-Dataset Evidence on Data Minimization, Output Restriction, and Differential Privacy.”

It contains the executable Google Colab notebook, prespecified scientific protocol, threat model, execution guide, dataset manifest, software requirements, the exact frozen PUBLICATION-mode result archive, a manuscript-to-results audit, and the secondary H4 monotonicity analysis.

The study evaluates membership leakage across four model families, four black-box attacks, two feature views, three release interfaces, and differentially private logistic regression on OULAD and UCI-697. The frozen outputs were checked against the manuscript values to displayed precision.

Raw source datasets are not redistributed. The notebook retrieves the public datasets from their original sources and validates the expected data configuration. Record-level exported membership scores contain no original student identifiers.