Research map / Adversarial robustness and security / Attacks on text and graphs

TAPAS: Datasets for Learning the Learning with Errors Problem

NeurIPS 2025 · Adversarial robustness and security › Attacks on text and graphs · approach: Attacks and defenses in sequential settings · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Releases five large-scale preprocessed (lattice-reduced) Learning with Errors datasets (40M–400M samples) on Hugging Face and establishes SALSA and Cool & Cruel attack baselines on them.

Problem

AI attacks on Learning with Errors (LWE) need heavy, expert preprocessing via lattice reduction, and existing public LWE datasets are small and cover few parameter regimes, which blocks AI researchers from entering the area.

Approach

Provide large-scale, off-the-shelf datasets of lattice-reduced LWE samples across several (n, q) settings generated with interleaved BKZ2.0/flatter reduction, plus baseline results of the two existing AI attacks (SALSA and Cool & Cruel) on them.

Key results

• SALSA recovers binary secrets up to Hamming weight 63 on n=512, log2 q=41, the highest across settings.
• Cool and Cruel recovers binary secrets up to Hamming weight 60 on n=512, log2 q=41 and 22 on n=256, log2 q=20, outperforming SALSA on some settings.
• Datasets total up to 400M samples for n=256 and 40M for other settings, roughly 10x-100x prior public LWE datasets (about 4M examples).

Surprising findings

• Neither SALSA nor Cool and Cruel dominates: the better attack varies by parameter setting and secret type.
• Recovery of secrets is limited to low Hamming weights on most settings (e.g., 5-6 for n=1024 and n=512, log2 q=12).

Limitations

• Reduction quality is only on par with prior datasets; improving it requires new lattice reduction methods outside the paper's scope.
• Dataset generation required substantial CPU compute (up to ~40M CPU hours).

Experimental setup

Datasets: TAPAS (five reduced LWE datasets, n=256/512/1024)

Benchmarks: TAPAS

Metrics: max Hamming weight of recovered binary secret max Hamming weight of recovered ternary secret stddev reduction factor rho number of cruel bits CPU hours per matrix

Baselines: SALSA Cool and Cruel

Builds on: SALSA Cool and Cruel BKZ 2.0 flatter LWE-benchmarking

Closest papers by meaning

• Making Hard Problems Easier with Custom Data Distributions and Loss Regularization: A Case Study in Modular Arithmetic (ICML 2025 · similarity 0.90)
• QCircuitBench: A Large-Scale Dataset for Benchmarking Quantum Algorithm Design (NeurIPS 2025 · similarity 0.84)
• Rethink the Role of Deep Learning towards Large-scale Quantum Systems (ICML 2025 · similarity 0.83)
• Mind the Gap: A Practical Attack on GGUF Quantization (ICML 2025 · similarity 0.83)
• Qsco: A Quantum Scoring Module for Open-Set Supervised Anomaly Detection (AAAI 2025 · similarity 0.83)
• Generating Universal Adversarial Perturbations for Quantum Classifiers (AAAI 2024 · similarity 0.82)