XposedMIA: Exploiting Explainable Artificial Intelligence (XAI) for One-Shot Membership Inference Attacks
Organizations: Bilkent University, Ankara, Turkey · Case Western Reserve University, Cleveland, Ohio, USA · IBM Research, Dublin, Ireland
Abstract
Machine learning (ML) services increasingly expose explainability outputs alongside model predictions. While such explanations are intended to improve transparency, they may also reveal information about the underlying training data. In this work, we show that the local surrogate geometry employed by Local Interpretable Model-agnostic Explanations (LIME) can be repurposed as a proxy for a sample's distance to the model's decision boundary. We extend this boundary-distance view to tabular counterfactual explanations and, for the first time, to visual counterfactual explanations for image classifiers. We present XposedMIA, a one-shot, explanation-based black-box membership inference attack (MIA) framework that converts the returned explanation into a proxy for the queried sample's distance to the model decision boundary. After a one-time calibration step on a small labeled calibration set of as few as 30 samples, our attack requires only a single query for each target sample. Unlike prior label-only attacks requiring iterative queries or shadow training, our method relies solely on the predicted label and its accompanying explanation. Our evaluation indicates that XposedMIA achieves high effectiveness, with success rates of up to 0.95 using only a single query per target, revealing that releasing explanations introduces an additional channel for membership leakage.
Figures & tables
| Dataset | Dataset Size | # Features |
| Purchase-100 | 197,324 | 600 |
| Texas-100 | 67,330 | 6,170 |
| Location | 5,010 | 446 |
| Adult | 48,842 | 14 |
| Dataset | Generator | ASR | ROC–AUC | CF Success | |
| M | N | ||||
| Purchase-100 | NICE | 84.8% | 91.6% | 100.0% | 100.0% |
| Growing Spheres | 95.0% | 97.6% | 88.7% | 98.9% | |
| Wachter | 92.1% | 96.7% | 99.6% | 100.0% | |
| Texas-100 | NICE | 62.7% | 67.6% | 100.0% | 100.0% |
| Growing Spheres | 82.0% | 88.1% | 78.4% | 95.9% | |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Dataset (distance) | Threshold: Gap Joint | ASR Gap (%) | ASR Joint (%) | ASR (pp) | |
| Image attacks | ||||||
| SEDC | CIFAR-10 (segment count) | 72.17 | 72.17 | + 0.00 | 9.95 9.95 | |
| SEDC | CIFAR-100 (segment count) | 65.54 | 68.42 | 9.70 52.96 | ||
| DiVE | CIFAR-10 (LPIPS) | 84.27 | 85.87 | 6.47 11.93 | ||
| DiVE | CIFAR-100 (SSIM) | 75.33 | 75.08 | 2.02 11.63 | ||
| Tabular attacks | ||||||
| Method | Attack signal | Experimental setup | Queries | Purchase | Location | Texas | ||||
| One-time | Per candidate | ASR | AUC | ASR | AUC | ASR | AUC | |||
| Gap Attack ( Yeom et al., 2018 ) | Prediction correctness | Evaluated on our target models; no calibration | 0 | 1 label | .692 | – | .661 | – | .652 | – |
| Choquette-Choo et al. ( Choquette-Choo et al., 2021 ) | Robustness to input perturbations | Shadow/source model for threshold selection; hard-label access | 0 target | labels | .874 | – | .892 | – | .803 | – |
| Shokri et al. ( Shokri et al., 2021 ) | Variance of LIME coefficients | Optimal variance threshold; Purchase and Texas only | – | 1 expl. a | – | – | – | – | ||
| Ye et al. ( Ye et al., 2022 ) (Attack P) | Population-calibrated loss score | 1,000 population records per class; score access | scores b | 1 score | – | .818 | – | – | – | – |
| YOQO ( Wu et al., 2024 ) | Shadow-crafted membership query | Online: 16 in/out pairs; offline: 16 out-models | 0 target | 1 label | – | – | – | – | ||
| Attack family | Dataset | Target model and member source | Method | per group | Attack-subset construction | |
| Tabular attacks | ||||||
| LIME | Purchase-100 | 4,800 / 1,200 fit/validation | LIME | 1,000 | 300 / 700 | Balanced |
| LIME | Location | 960 / 240 fit/validation | LIME | 900 | 270 / 630 | Balanced |
| LIME | Texas-100 | 8,000 / 2,000 fit/validation | LIME | 1,000 | 300 / 700 | Not class-balanced |
| Counterfactual | Purchase-100 | 8,000 / 2,000 fit/validation | NICE, Wachter, Growing Spheres | 1,000 | 300 / 700 | Not class-balanced |
| Counterfactual | Location | 960 / 240 fit/validation | NICE, Wachter, Growing Spheres | 1,200 | 270 / 630 | Balanced |
Membership Inference Attacks
Membership inference attacks test whether records or entities were used to train a machine-learning model, revealing potential privacy leakage.
See all 124 papers