Research map / Adversarial robustness and security / Certified and empirical robustness
Do Perceptually Aligned Gradients Imply Robustness?
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 score-based gradient PAG-inducing auxiliary loss (no adversarial training) | ResNet-18 standard training · CIFAR-10 | AA L2 (eps=0.5) accuracy: 0.00% -> 55.39% | Helped |
| add score-based gradient PAG-inducing auxiliary loss (no adversarial training) | ResNet-18 standard training · CIFAR-10 | clean accuracy: 93.61% -> 78.56% | Hurt |
| add score-based gradient PAG-inducing auxiliary loss | ViT trained from scratch · CIFAR-10 | AA L2 (eps=0.5) accuracy: 42.99% -> 57.80% over AT L2 | Helped |
| add score-based gradient PAG-inducing auxiliary loss | ViT trained from scratch · CIFAR-10 | clean accuracy: 80.51% -> 81.28% | Helped |
| add score-based gradient PAG-inducing auxiliary loss | ResNet-18 on STL in low-data regime · STL | AA L2 (eps=0.5) accuracy: 46.04% (AT L2) -> 65.96% | Helped |
| add score-based gradient PAG-inducing auxiliary loss | ResNet-18 on CIFAR-100 · CIFAR-100 | AA L2 (eps=0.5) accuracy: 29.25% vs 30.51% for AT L2 | No clear effect |
| add score-based gradient PAG-inducing auxiliary loss | ResNet-18 evaluated on large-eps attacks · CIFAR-10 | AA L2 (eps=1.0) accuracy: 18.81% (AT L2) -> 30.43% | Helped |
| add gradient alignment (PAG) loss to AT L2 (ablation) | AT L2 on ResNet-18 · CIFAR-10 | AA L2 (eps=0.5) accuracy: 60.82% -> 61.73% | Helped |
| add gradient alignment (PAG) loss to AT L2 (ablation) | AT L2 on ResNet-18 · CIFAR-10 | AA Linf (eps=8/255) unseen attack accuracy: 19.63% -> 23.52% | Helped |
| add gradient alignment (PAG) loss to AT L2 (ablation) | AT L2 on ResNet-18 · CIFAR-10 | clean accuracy: 86.79% -> 85.54% | Hurt |
| add gradient alignment (PAG) loss to TRADES Linf (ablation) | TRADES on WRN 34-10 · CIFAR-10 | AA Linf (eps=8/255) accuracy: 43.54% -> 45.78% | Helped |
| increase PAG regularization weight lambda (0 to 2.0) in SBG (ablation) | ResNet-18 SBG · CIFAR-10 | robustness (qualitative PAG and robustness) | Helped |
Closest papers by meaning
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- Can Implicit Bias Imply Adversarial Robustness? (ICML 2024 · similarity 0.88)
- The Implicit Bias of Gradient Descent toward Collaboration between Layers: A Dynamic Analysis of Multilayer Perceptions (NeurIPS 2024 · similarity 0.88)
- Attack-free Evaluating and Enhancing Adversarial Robustness on Categorical Data (ICML 2024 · similarity 0.88)
- Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks (ICLR 2025 · similarity 0.88)
- On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness (ICML 2025 · similarity 0.88)
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.