Research map / Adversarial robustness and security / Certified and empirical robustness
Reliably Fast Adversarial Training via Latent Adversarial Perturbation
Geon Yeong Park, Sang Wan Lee
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 latent adversarial perturbation (SLAT) to single-step FGSM AT (ablation) | FGSM AT · CIFAR-10 | PGD-50-10 robust accuracy (%): 0.01 -> 47.06 | Improved |
| add latent adversarial perturbation to FGSM-RS (ablation) | FGSM-RS · CIFAR-10 | PGD-50-10 robust accuracy (%): 3.90 -> 44.95 | Improved |
| derive latent perturbation from the original sample rather than the randomly perturbed one (SLAT vs FGSM-RS + latent) (ablation) | SLAT · CIFAR-10 | PGD-50-10 robust accuracy (%): 44.95 -> 47.06 | Improved |
| reduce FGSM step size to 0.9*8/255 instead of using latent perturbation (ablation) | FGSM AT · CIFAR-10 | PGD-50-10 robust accuracy (%): 0.01 -> 3.09 | No clear effect |
| inject latent perturbation into conv3, conv4 instead of conv1, conv2 (ablation) | SLAT · CIFAR-10 | PGD-50-10 robust accuracy (%): 47.06 -> 0.00 | Hurt |
| inject latent perturbation into all of conv1-conv4 in addition to input (ablation) | SLAT · CIFAR-10 | standard accuracy (%): 85.91 -> 81.41 | Hurt |
| inject latent perturbation into all of conv1-conv4 in addition to input (ablation) | SLAT · CIFAR-10 | AutoAttack robust accuracy (%): 44.62 -> 43.58 | Hurt |
| inject latent perturbation into all of conv1-conv4 in addition to input (ablation) | SLAT · CIFAR-10 | PGD-50-10 robust accuracy (%): 47.06 -> 47.3 | No clear effect |
| replace SLAT with FGSM-CKPT | SLAT vs FGSM-CKPT · CIFAR-10 | PGD-50-10 robust accuracy (%): 40.83 -> 47.06 | Improved |
| compare SLAT with Free-AT (m=8) | SLAT vs Free-AT · CIFAR-10 | PGD-50-10 robust accuracy (%): 44.15 -> 47.06 | Improved |
| compare SLAT with PGD-7 AT | SLAT vs PGD-7 · CIFAR-10 | training time (min): 383.2 -> 104.6 | Improved |
| compare SLAT with FGSM-GA | SLAT vs FGSM-GA · CIFAR-10 | training time (min): 297.9 -> 104.6 | Improved |
Closest papers by meaning
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.90)
- SORA: Free Second-Order Attacks in Fast Adversarial Training (ICML 2026 · similarity 0.88)
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · similarity 0.88)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · 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.
