DeepLeak
Jan 1, 2026·
·
1 min read
Firas Ben Hmida
Zain Sbeih
Philemon Hailemariam
Birhanu Eshete
Project Snapshot
| Item | Details |
|---|---|
| Paper | DeepLeak: Privacy Enhancing Hardening of Model Explanations Against Membership Leakage |
| Venue | IEEE SaTML 2026 |
| Primary Theme | Privacy-preserving explainability |
| Main Artifacts | Codebase + dataset package |
Authors
- Firas Ben Hmida
- Zain Sbeih
- Philemon Hailemariam
- Birhanu Eshete
Overview
DeepLeak studies the privacy risks of post-hoc explanation methods and provides mitigation strategies that reduce membership leakage while preserving explanation utility. The project focuses on practical, model-agnostic hardening that can be applied in high-stakes ML deployments.
What This Project Delivers
- Explanation-aware leakage auditing across multiple explanation families.
- Hardening strategies including attribution clipping, masking, and calibrated noise.
- Reproducible artifacts to evaluate privacy/utility tradeoffs under consistent settings.
Repository and Paper
- Code: github.com/um-dsp/DeepLeak
- Datasets: github.com/um-dsp/DeepLeak/datasets
- Paper: arxiv.org/abs/2601.03429v1
