Research map / Adversarial robustness and security / Certified and empirical robustness
Guided Adversarial Attack for Evaluating and Enhancing Adversarial Defenses
Gaurang Sriramanan, Sravanti Addepalli, Arya Baburaj, Venkatesh Babu R
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 |
|---|---|---|---|
| use GAT instead of FBF single-step defense | RN18 single-step adversarial training · CIFAR-10 | AutoAttack accuracy (%): 43.12 -> 47.30 | Improved |
| add GAMA adversaries to TRADES (TR-GAT) | TRADES WRN34 · CIFAR-10 | AutoAttack accuracy (%): 53.18 -> 53.32 | Improved |
| use margin loss instead of cross-entropy loss in PGD attack (ablation) | PGD-100 · TRADES-WRN34 CIFAR-10 | accuracy under attack (%): 55.61 -> 53.94 | Improved |
| compute margin on probabilities instead of logits (ablation) | PGD margin attack · TRADES-WRN34 CIFAR-10 | accuracy under attack (%): 54.19 -> 53.94 | Improved |
| add decaying L2 relaxation term (graduated optimization) with Bernoulli init (GAMA-PGD) (ablation) | PGD margin attack in prob. space · TRADES-WRN34 CIFAR-10 100-step | accuracy under attack (%): 53.94 -> 53.29 | Improved |
| use Frank-Wolfe (GAMA-FW) instead of PGD for 10-step attack (ablation) | GAMA · TRADES-WRN34 CIFAR-10 10-step | accuracy under attack (%): 54.95 -> 54.27 | Improved |
| use Adam with true gradients instead of signed SGD in GAMA-PGD (ablation) | GAMA-PGD · TRADES-WRN34 CIFAR-10 | accuracy under attack (%): 53.29 -> 53.66 | Hurt |
| remove alternation between CE and GA-CE attacks in GAT (ablation) | GAT · CIFAR-10 RN18 | AutoAttack accuracy (%): 46.52 -> 47.30 | Improved |
| use GAMA max-margin loss for training in GAT instead of cross-entropy (ablation) | GAT · CIFAR-10 RN18 | AutoAttack accuracy (%): 47.30 -> 10.98 | Hurt |
| use both guided attack and guided defense loss instead of only guided attack (A3) or only defense (A5) (ablation) | GAT · CIFAR-10 RN18 | AutoAttack accuracy (%): A3 32.29, A5 44.21 -> 47.30 | Improved |
Closest papers by meaning
- GAT: Guided Adversarial Training with Pareto-optimal Auxiliary Tasks (ICML 2023 · similarity 0.87)
- Improving Adversarial Robustness via Guided Complement Entropy (ICCV 2019 · similarity 0.87)
- Variational Adversarial Defense: A Bayes Perspective for Adversarial Training (TPAMI 2023 · similarity 0.87)
- GAMA: Generative Adversarial Multi-Object Scene Attacks (NeurIPS 2022 · similarity 0.86)
- Regional Adversarial Training for Better Robust Generalization (IJCV 2024 · similarity 0.86)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.86)
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.
