Research map / Adversarial robustness and security / Certified and empirical robustness
Characterizing Model Robustness via Natural Input Gradients
Adrian Rodriguez-Munoz, Tongzhou Wang, Antonio Torralba
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 |
|---|---|---|---|
| regularize input gradient norm on natural examples only instead of adversarial training | Swin-B · ImageNet-1k AutoAttack eps=4/255 | AutoAttack accuracy: 0.00 -> 51.58 (adversarial training 56.12) | Improved |
| replace ReLU with smooth GeLU in ResNet-50 for gradient norm regularization (ablation) | ResNet-50 GradNorm · ImageNet-1k AutoAttack | AutoAttack accuracy: 6.82 -> 30.00 | Improved |
| replace ReLU with GeLU in ResNet-50 for adversarial training (ablation) | ResNet-50 adversarial training · ImageNet-1k AutoAttack | AutoAttack accuracy: 31.60 -> 32.64 | Improved |
| increase GradNorm weight from 1.2 to 1.4 (ablation) | Swin-B GradNorm · ImageNet-1k AutoAttack eps=4/255 | AutoAttack accuracy: 51.58 -> 52.48 AA; clean 77.78 -> 76.20 | Mixed |
| double the GradNorm weight on the green channel (ablation) | Swin-B GradNorm · ImageNet-1k | AutoAttack accuracy: 51.66 -> 51.38 | Hurt |
| add edge regularization aligning class gradients with image edges | Swin-B · ImageNet-1k AutoAttack eps=4/255 | AutoAttack accuracy: 0.00 -> 35.02 | Improved |
| regularize saliency maps or loss gradients toward edges instead of class gradients (ablation) | Edge regularization · ResNet50+GeLU ImageNet | robustness | Hurt |
| add adversarial-noise inputs to gradient norm regularization (ablation) | Swin-B GradNorm · ImageNet-1k AutoAttack eps=4/255 | AutoAttack accuracy: 51.66 -> 56.49 | Improved |
| add Gaussian noise inputs to gradient norm regularization (ablation) | Swin-B GradNorm · ImageNet-1k AutoAttack eps=4/255 | AutoAttack accuracy: 51.66 -> 52.04 | Improved |
| gradient norm regularization instead of adversarial training on CIFAR-10 | PreActResNet-18 · CIFAR-10 eps=8/255 | AutoAttack accuracy: AA 42.05 -> 39.60; clean 69.81 -> 77.74 | Mixed |
Closest papers by meaning
- Adversarial Robustness through Local Linearization (NeurIPS 2019 · similarity 0.92)
- (Certified!!) Adversarial Robustness for Free! (ICLR 2023 · similarity 0.91)
- Towards Understanding the Regularization of Adversarial Robustness on Neural Networks (ICML 2020 · similarity 0.91)
- What It Thinks Is Important Is Important: Robustness Transfers Through Input Gradients (CVPR 2020 · similarity 0.91)
- Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients (AAAI 2018 · similarity 0.91)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
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.
