Accepted to the AAAI’26 Workshop on Shaping Responsible Synthetic Data in the Era of Foundation Models (RSD).
MIA-EPT is a novel black-box membership inference attack designed for tabular diffusion models. It constructs error-based feature vectors by masking and reconstructing attributes of target records, and infers membership from how well attributes can be predicted using only the released synthetic data.
We validate MIA-EPT across multiple diffusion-based synthesizers, achieving AUC-ROC up to 0.599 and TPR@10% FPR up to 22.0% in internal tests, and 2nd place in the MIDST 2025 Black-box Multi-Table track (TPR@10% FPR = 20.0%).
MIA-EPT follows a shadow-model pipeline: train shadow diffusion models, train per-column attribute predictors on synthetic data, extract structured error profiles for records, and train an attack classifier to predict membership.
We evaluate MIA-EPT on diffusion-based tabular synthesizers (including single-table and multi-table settings) and report AUC-ROC and TPR@10% FPR, matching the MIDST competition operating point.
| Setting | Metric | Value |
|---|---|---|
| Internal evaluation | Best AUC-ROC | 0.599 |
| Internal evaluation | Best TPR@10% FPR | 22.0% |
| MIDST 2025 (Black-box Multi-Table) | TPR@10% FPR | 20.0% (2nd place) |
MIA-EPT achieved 2nd place in the MIDST 2025 Black-box Multi-Table track (TPR@10% FPR = 20.0%), demonstrating measurable membership leakage in diffusion-based synthetic tabular data.
@misc{german2025mia_ept,
title = {MIA-EPT: Membership Inference Attack via Error Prediction for Tabular Data},
author = {Eyal German and Daniel Samira and Yuval Elovici and Asaf Shabtai},
year = {2025},
eprint = {2509.13046},
archivePrefix= {arXiv},
primaryClass = {cs.CR},
url = {https://arxiv.org/abs/2509.13046}
}