Research map / Adversarial robustness and security / Certified and empirical robustness
Verifying Neural Network Robustness with Dual Perturbations
Hai Duong, Lam Nguyen, Thanh Le, ThanhVu Nguyen
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 independent pixel perturbation to universal convolutional perturbation (dual) | VeriDou · MNIST-FC | SAT instances (%): 12% -> 75% | Improved |
| use universal instead of restricted convolutional kernels (ablation) | VeriDou · TinyImageNet | SAT instances (%): 53% -> 89% | Improved |
| use restricted convolutional perturbation instead of independent only (ablation) | VeriDou · TinyImageNet | SAT instances (%): 20% -> 53% | Improved |
| add convolutional component (zij from 0.0 to 0.1) to independent perturbation (ablation) | VeriDou · CIFAR100 | SAT instances (%): 0% -> 85% | Improved |
| add convolutional component (zij from 0.0 to 0.1) to independent perturbation (ablation) | VeriDou · TinyImageNet | SAT instances (%): 20% -> 70% | Improved |
| increase independent perturbation radius from 0.01 to 0.04 (ablation) | VeriDou · MNIST-FC | SAT instances (%): 21% -> 33% | Improved |
| increase independent perturbation radius from 0.01 to 0.04 (ablation) | VeriDou · Oval21 | SAT instances (%): 15% -> 20% | Improved |
| add independent noise to universal kernels on continuous angle ranges (ablation) | VeriDou alpha-beta · Continuous kernel angle ranges | SAT count: 142 -> 171 | Improved |
| increase kernel entry coverage from 20% to 100% for arbitrary kernels (ablation) | VeriDou NeuralSAT · Arbitrary kernels (Domain 3) | timeout instances: 75 -> 165 | Hurt |
| increase number of evaluated samples from 300 to 600 (ablation) | VeriDou NeuralSAT · Motion 0 degrees | SAT fraction (dual C=1.0): 0.68 -> 0.76 | No clear effect |
| switch between verifier backends or perturbation types on runtime | VeriDou · runtime comparison | runtime | No clear effect |
Closest papers by meaning
- Evaluating Robustness of Neural Networks with Mixed Integer Programming (ICLR 2019 · similarity 0.90)
- Verification of Neural Networks Against Convolutional Perturbations via Parameterised Kernels (AAAI 2025 · similarity 0.89)
- Training for Faster Adversarial Robustness Verification via Inducing ReLU Stability (ICLR 2019 · similarity 0.89)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.88)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.87)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.87)
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.
