Research map / Adversarial robustness and security / Certified and empirical robustness
Which Models have Perceptually-Aligned Gradients? An Explanation via Off-Manifold Robustness
Suraj Srinivas, Sebastian Bordt, Himabindu Lakkaraju
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 |
|---|---|---|---|
| increase robustness regularization (gradient norm, randomized smoothing, smoothness penalty) from none to moderate | Resnet18 · CIFAR-10 | off-manifold sensitivity | Improved |
| increase robustness regularization from none to moderate | Resnet18 · CIFAR-10 | on-manifold sensitivity | No clear effect |
| increase robustness regularization beyond the intermediate regime (excessive robustness) | Resnet18 · CIFAR-10 | accuracy | Hurt |
| increase robustness regularization from weak to intermediate | Resnet18 input gradients · CIFAR-10 | LPIPS perceptual similarity to score | Improved |
| increase robustness regularization beyond the peak | Resnet18 input gradients · CIFAR-10 | LPIPS perceptual similarity to score | Hurt |
| train excessively robust PGD models | Resnet18 · ImageNet-64x64 | accuracy and perceptual alignment | Hurt |
| use l2-adversarially robust training instead of standard training | Resnet18 · MNIST with distractor | relative noise robustness (distractor vs signal) | Improved |
| use adversarially robust pretrained models | Resnet18 · ImageNet | LPIPS perceptual similarity to score | Improved |
Closest papers by meaning
- Do Perceptually Aligned Gradients Imply Robustness? (ICML 2023 · similarity 0.92)
- Generative Model Inversion Through the Lens of the Manifold Hypothesis (NeurIPS 2025 · similarity 0.87)
- Characterizing Model Robustness via Natural Input Gradients (ECCV 2024 · similarity 0.87)
- The Geometry of Robustness: Optimizing Loss Landscape Curvature and Feature Manifold Alignment for Robust Finetuning of Vision-Language Models (CVPR 2026 · similarity 0.87)
- Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision (NeurIPS 2026 · similarity 0.85)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · 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.
