Research map / Adversarial robustness and security / Attacks on text and graphs
Fundamental Limits of Membership Inference Attacks on Machine Learning Models
Eric Aubinais, Elisabeth Gassiat, Pablo Piantanida
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| discretize the data (cluster) before training, C_K(P)=4.30 | 3-layer network · MNIST (n=1000) | accuracy: 0.989 -> 0.963 | Hurt |
| discretize the data with C_K(P)=6.74 or 9.20 | 3-layer network · MNIST (n=1000) | accuracy: 0.989 -> 0.968 and 0.986 | Hurt |
| use a coarser discretization (smaller C_K(P), 4.3 vs 9.20) (ablation) | 3-layer network · MNIST (n=1000) | security gap bound Delta: 0.15 -> 0.07 | Improved |
| use a coarser discretization (C_K(P)=4.3 vs 9.20) (ablation) | 3-layer network · MNIST (n=1000) | accuracy: 0.986 -> 0.963 | Hurt |
| increase dataset size n from 1000 to 10000 (ablation) | discretized MNIST, C_K(P)=4.3 · MNIST | security gap bound Delta: 0.07 -> 0.02 | Improved |
| increase dataset size n from 1000 to 10000 (ablation) | raw MNIST network · MNIST | accuracy: 0.989 -> 0.994 | Improved |
| train to overfitting (2500 iterations) then apply loss-threshold attack | 2-layer ReLU network, regression · non-linear regression | training vs validation accuracy at eps=1e-6: 100% training vs near 0% validation | Hurt |
| increase input dimension from 10 to 2000 (ablation) | overfitted linear regression · linear regression | validation accuracy at eps=1e-6 | No clear effect |
Closest papers by meaning
- Low-Cost High-Power Membership Inference Attacks (ICML 2024 · similarity 0.90)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 2025 · similarity 0.90)
- Confidence Is All You Need for MI Attacks (Student Abstract) (AAAI 2024 · similarity 0.90)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.89)
- Exploring the limits of strong membership inference attacks on large language models (NeurIPS 2025 · similarity 0.89)
- Causal Evaluation of Membership Inference Attacks (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
