Research map / Adversarial robustness and security / Attacks on text and graphs
SALSA: Attacking Lattice Cryptography with Transformers
Emily Wenger, Mingjie Chen, Francois Charton, Kristin 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 |
|---|---|---|---|
| replace Universal Transformer with regular 8-layer transformer (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 22.5 -> 26.3 | Hurt |
| add copy-gate mechanism (ablation) | SALSA UT · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 26.5 -> 22.6 | Improved |
| iterate shared layers 2/8 times instead of 4/4 (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 26.1 -> 23.5 | Improved |
| increase encoder dimension from 1024 to 3040 (decoder fixed at 512) (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 22.5 -> 19.7 | Improved |
| increase decoder dimension from 1024 to 1536 (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 23.9 -> 24.3 | Hurt |
| use small batch size (<50) (ablation) | SALSA · SALSA n=50, Hamming weight 3 | samples needed | Improved |
| reuse samples via linear combinations (K and times reused increase) (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 20.42 -> 15.405 | Improved |
| use higher input base (81 instead of 7) (ablation) | SALSA · SALSA n=50, Hamming weight 3 | log2 samples to recover secret: 24.0 -> 22.1 | Improved |
| bound the coordinate range of a (a_i < alpha q) (ablation) | SALSA · n=50 secret density | max recoverable density: 0.08 -> 0.3 | Improved |
| increase error sigma up to 24 (ablation) | SALSA · SALSA n=50, Hamming weight 3 | samples needed | Hurt |
| use distinguisher recovery instead of direct recovery only (ablation) | SALSA · secret recovery | share of recoveries: 55% direct, 18% distinguisher, 27% both | Mixed |
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)
- TAPAS: Datasets for Learning the Learning with Errors Problem (NeurIPS 2025 · similarity 0.87)
- BiShield-TEE: On the (In-)Security of Unilateral Weight Obfuscation in On-Device TEE-Shielded LLM Partition (NeurIPS 2026 · similarity 0.79)
- Unelicitable Backdoors via Cryptographic Transformer Circuits (NeurIPS 2024 · similarity 0.79)
- Exact Soft Analytical Side-Channel Attacks using Tractable Circuits (ICML 2024 · similarity 0.78)
- Cryptographic Hardness of Score Estimation (NeurIPS 2024 · similarity 0.78)
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.
