Research map / Adversarial robustness and security / Certified and empirical robustness
Improving ML Attacks on LWE with Data Repetition and Stepwise Regression
Alberto Alfarano, Eshika Saxena, Emily Wenger, Francois Charton, Kristin E. Lauter
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 |
|---|---|---|---|
| add data repetition and larger training sets | transformer LWE attack · n=64 log2 q=20 unreduced, h=4-6 | secret recovery | Improved |
| add repetition (small data budget) (ablation) | transformer LWE attack · n=256 log2 q=12 BKZ, 4M samples | highest h/cruel bits: 10/3 -> 14/5 | Improved |
| add repetition (15x vs 1x) (ablation) | transformer LWE attack · n=256 log2 q=20 BKZ, 4M samples | highest h/cruel bits: 55/6 -> 60/8 | Improved |
| replace linear regression with stepwise regression for cool bits (ablation) | cool bit recovery · n=256 log2 q=20, 38 cool bits, 20M | secrets recovered out of 20: 1 -> 11 | Improved |
| add dual stepwise regression on top of stepwise (ablation) | cool bit recovery · n=256 log2 q=20, 45 cool bits, 20M | secrets recovered out of 20: 3 -> 18 | Improved |
| use dual stepwise regression instead of linear regression (ablation) | cool bit recovery · n=512 log2 q=41, 80 cool bits, 1M | secrets recovered out of 20: 0 -> 19 | Improved |
| use LASSO or Ridge instead of linear regression (ablation) | cool bit recovery · cool bit recovery | secrets recovered | No clear effect |
| add cool/cruel region embedding (ablation) | transformer input · n=256 log2 q=20, 400M samples | secrets with 4 cruel bits recovered: 2/5 -> 5/5 | Improved |
| train on synthetic instead of BKZ-reduced data (ablation) | transformer LWE attack · all four settings | highest h/cruel bits | No clear effect |
| increase model size (embedding 256-1024, 4-12 layers) (ablation) | transformer LWE attack · n=256 log2 q=20, h=60/65/70, 100M data | model-based attempts A | No clear effect |
| use penalty alpha=0.1 beta=0.1 instead of alpha=0 beta=1 (ablation) | loss function · n=256 log2 q=20, 400M 2x | highest h/cruel bits: 65/8 -> 70/8 | Improved |
| add repetition (ablation) | transformer LWE attack · n=256 log2 q=20 | scaling exponent alpha_R: 0.70 -> 1.31 | Improved |
Closest papers by meaning
- Making Hard Problems Easier with Custom Data Distributions and Loss Regularization: A Case Study in Modular Arithmetic (ICML 2025 · similarity 0.89)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.86)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.86)
- TAPAS: Datasets for Learning the Learning with Errors Problem (NeurIPS 2025 · similarity 0.86)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.85)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.85)
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.
