Research map / Adversarial robustness and security / Certified and empirical robustness
Evaluating Robustness of Neural Networks with Mixed Integer Programming
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 MIPVerify rather than PGD to find adversarial examples | LPd-CNNa · MNIST, l∞ ε=0.1 | adversarial error lower bound: 4.11% -> 4.38% | Helped |
| Use MIPVerify rather than the prior state-of-the-art certificate | LPd-CNNa · MNIST, l∞ ε=0.1 | adversarial error upper bound: 5.82% -> 4.38% | Helped |
| Use progressive bounds tightening (ablation) | MIPVerify · MNIST, LPd-CNNa, l∞ ε=0.1 | total verification time: 7.77 -> 3.52 | Helped |
| Use bounds from the restricted perturbation domain (ablation) | MIPVerify · MNIST, LPd-CNNa, l∞ ε=0.1 | total verification time: 57.96 -> 3.52 | Helped |
| Use asymmetric bounds in the ReLU formulation (ablation) | MIPVerify · MNIST, LPd-CNNa, l∞ ε=0.1 | total verification time: 4598.15 -> 3.52 | Helped |
| Use MIPVerify rather than Reluplex | MNIST minimum targeted adversarial distortion · MNIST test samples, l∞ norm | verification speed: two to three orders of magnitude | Helped |
| Replace full-domain bounds with bounds over each sample's restricted perturbation domain | MIPVerify · Per-sample adversarial robustness verification | number of unstable ReLUs and labels considered | Helped |
| Use PGD instead of exact verification to identify vulnerable samples | LPd-CNNa classifiers · MNIST; minimum adversarial distortion approaches ε | attack success rate | Hurt |
Closest papers by meaning
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.90)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.90)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.90)
- Tightening Robustness Verification of MaxPool-based Neural Networks via Minimizing the Over-Approximation Zone (CVPR 2025 · similarity 0.90)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.89)
- Probably Approximately Global Robustness Certification (ICML 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.