Research map / Adversarial robustness and security / Certified and empirical robustness
A Convex Relaxation Barrier to Tight Robustness Verification of Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use exact optimal layer-wise convex relaxation (LP-all) instead of greedy LP-relaxed bounds | LP-relaxed verification · MNIST robust-error bounds, normally and adversarially trained networks | robust error upper bound: LP-all gives a tighter bound than LP-greedy in each case | Helped |
| Use exact optimal layer-wise convex relaxation (LP-all) instead of greedy LP-relaxed bounds | LP-relaxed verification · MNIST robust-error bounds, LPd-trained networks | robust error upper bound: 18.32% -> 17.24%; 41.67% -> 37.50%; 66.85% -> 60.17%; 87.89% -> 83.62% | Helped |
| Use optimal layer-wise convex relaxation rather than greedy relaxed bounds | LP-relaxed verification · MNIST minimum adversarial distortion | lower bound on distortion relative to PGD upper bound: 1.5 to 5 times smaller; below 1.5 times smaller for LPd-trained networks | Mixed |
| Use LP-relaxed verification rather than PGD to bound minimum adversarial distortion | LP-all, LP-last, and LP-greedy · CIFAR-10 minimum adversarial distortion | lower bound on distortion relative to PGD upper bound: 1.5 and 2 times smaller | Hurt |
Closest papers by meaning
- The Convex Relaxation Barrier, Revisited: Tightened Single-Neuron Relaxations for Neural Network Verification (NeurIPS 2020 · similarity 0.94)
- Convex Formulations for Training Two-Layer ReLU Neural Networks (ICLR 2025 · similarity 0.90)
- Expressivity of ReLU-Networks under Convex Relaxations (ICLR 2024 · similarity 0.90)
- Stress-Testing Neural Network Verifiers with Provably Robust Instances (NeurIPS 2026 · similarity 0.89)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.89)
- Verifying Properties of Binary Neural Networks Using Sparse Polynomial Optimization (ICLR 2025 · 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.