Research map / Adversarial robustness and security / Certified and empirical robustness
A Hamiltonian Monte Carlo Method for Probabilistic Adversarial Attack and Learning
Hongjun Wang, Guanbin Li, Xiaobai Liu, Liang Lin
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 accumulated momentum with sampling (HMCAM) to M-PGD | M-PGD · CIFAR10 black-box ResNetXt -> ResNet32_A | attack success rate: 10.11% -> 14.11% | Improved |
| use HMCAM instead of M-PGD for white-box attack on VGG16 | M-PGD · CIFAR10 white-box VGG16 | attack success rate: 83.64% -> 93.29% | Improved |
| use HMCAM instead of M-PGD on ResNet32 white-box | M-PGD · CIFAR10 white-box ResNet32 | attack success rate: 98.93% -> 98.76% | No clear effect |
| AI-FGSM (S=1, accumulated momentum only) vs full HMCAM with sampling S=2 (ablation) | AI-FGSM · CIFAR10 single-model black-box ResNet32 -> ResNet32_A | attack success rate: 9.56% -> 10.01% | Improved |
| add accumulated momentum (AI-FGSM) to M-PGD (ablation) | M-PGD · CIFAR10 ensemble hold-out ResNet32_A | attack success rate: 21.23% -> 28.31% | Improved |
| add HMCAM to ensemble attack on adversarially trained hold-out | PGD · CIFAR10 ensemble hold-out ResNet32_A | attack success rate: 13.88% -> 30.11% | Improved |
| use HMCAM with N=10 iterations versus PGD with N=100 (ablation) | PGD · CIFAR10 black-box | attack success rate | Improved |
| vary step size alpha over 0.001 to 0.1 times e^-2 (ablation) | HMCAM · CIFAR10 white/black-box | attack success rate | No clear effect |
| replace PGD-10 adversarial training with CAT on Preact-ResNet18 | PGD-10 adversarial training · CIFAR10 Preact-ResNet18 PGD-20 | robust accuracy: 41.58% -> 49.37% | Improved |
| replace PGD-10 adversarial training with CAT on Preact-ResNet18 | PGD-10 adversarial training · CIFAR10 Preact-ResNet18 clean | clean accuracy: 84.96% -> 81.54% | Hurt |
| replace PGD-10 with CAT on Wide ResNet34 | PGD-10 adversarial training · CIFAR10 WRN34 PGD-20 | robust accuracy: 48.4% -> 53.3% | Improved |
| replace PGD-10 with CAT on Wide ResNet34 (training time) | PGD-10 adversarial training · CIFAR10 WRN34 | training time (mins): 2602 -> 672 | Improved |
Closest papers by meaning
- Alleviating Adversarial Attacks on Variational Autoencoders with MCMC (NeurIPS 2022 · similarity 0.88)
- A-NICE-MC: Adversarial Training for MCMC (NeurIPS 2017 · similarity 0.87)
- Bayesian GAN (NeurIPS 2017 · similarity 0.87)
- Generalizing Hamiltonian Monte Carlo with Neural Networks (ICLR 2018 · similarity 0.87)
- Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning (CVPR 2024 · similarity 0.86)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.85)
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.
