Research map / Adversarial robustness and security / Certified and empirical robustness
Combating Adversaries with Anti-adversaries
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 |
|---|---|---|---|
| add Anti-Adv to robustly trained models | TRADES, ImageNet-Pre, MART, HYDRA, and AWP · CIFAR10 | AutoAttack accuracy: 53.11 -> 71.71; 55.31 -> 76.01; 56.75 -> 76.76; 57.64 -> 76.39; 60.53 -> 79.21 | Helped |
| add Anti-Adv to robustly trained models | ImageNet-Pre and AWP · CIFAR100 | AutoAttack accuracy: 28.96 -> 40.68; 29.15 -> 39.57 | Helped |
| add Anti-Adv to robustly trained models | TRADES, ImageNet-Pre, MART, HYDRA, and AWP · CIFAR10 | Square attack accuracy: 53.1 -> 71.7; 62.4 -> 78.5; 64.9 -> 78.0; 65.0 -> 78.8; 66.2 -> 80.7 | Helped |
| add Anti-Adv to robustly trained models | ImageNet-Pre and AWP · CIFAR100 | Square attack accuracy: 34.6 -> 42.4; 34.7 -> 46.4 | Helped |
| add Anti-Adv | nominally trained ResNet18 · CIFAR10, NES, 5K queries | accuracy: 5.8 -> 77.0 | Helped |
| replace Anti-Adv with SND | nominally trained ResNet50 · ImageNet, NES, 5K queries | accuracy: 67.2 -> 65.4 | Hurt |
| add Anti-Adv | AWP · CIFAR10 | clean accuracy: 88.5 -> 88.5 | No clear effect |
| add Anti-Adv to AWP | AWP · CIFAR10, Boundary Attack and QEBA | accuracy: 86.0 -> 86.0 (BA); 85.0 -> 86.0 (QEBA) | Mixed |
| vary the anti-adversary learning rate alpha across the tested values (ablation) | AWP + Anti-Adv · CIFAR10, Square and AutoAttack | robust accuracy | No clear effect |
| increase solver iterations K from 2 to 3 (ablation) | AWP + Anti-Adv · CIFAR10 | robust accuracy | Helped |
| use K=1 instead of the usual K=2 (ablation) | robustly trained models with Anti-Adv · CIFAR10 and CIFAR100 | AutoAttack accuracy: improvement against AutoAttack is 13% on CIFAR10 and 5% on CIFAR100 | Helped |
| increase solver iterations K (ablation) | AWP + Anti-Adv · CIFAR10 | robust accuracy | Helped |
Closest papers by meaning
- On Detecting Adversarial Perturbations (ICLR 2017 · similarity 0.90)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.89)
- Deep Defense: Training DNNs with Improved Adversarial Robustness (NeurIPS 2018 · similarity 0.89)
- Combining Adversaries with Anti-adversaries in Training (AAAI 2023 · similarity 0.89)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.88)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · 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.