Research map / Adversarial robustness and security / Certified and empirical robustness
What It Thinks Is Important Is Important: Robustness Transfers Through Input Gradients
Check 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 |
|---|---|---|---|
| transfer robustness through input-gradient matching from an MNIST robust teacher | CIFAR-10 classifier · CIFAR-10 test, PGD20 | accuracy (%): 0 -> 43.5 | Helped |
| replace frozen-feature logit finetuning with input-gradient matching from a Tiny-ImageNet teacher | CIFAR-10 classifier · CIFAR-10 test, PGD20 | accuracy (%): 24.9 -> 45.1 | Helped |
| use input resizing rather than random padding for teacher input adaptation | IGAM with a Tiny-ImageNet teacher · CIFAR-10 test, PGD20 | accuracy (%): 38.6 -> 45.1 | Helped |
| transfer robustness through input-gradient matching from a robust teacher | CIFAR-100 classifier · CIFAR-100 test, PGD20 | accuracy (%): 0 -> 21.74 | Helped |
| use a CIFAR-10 rather than MNIST robust teacher | IGAM on CIFAR-100 · CIFAR-100 test, PGD20 | accuracy (%): 17.67 vs 21.74 | Helped |
| add the gradient reconstruction loss (lambda_diff=5 rather than 0) (ablation) | IGAM-CIFAR10 · CIFAR-100 test | PGD20 accuracy (%): 16.0 -> 21.7 | Helped |
| increase gradient reconstruction-loss weight from 5 to 10 (ablation) | IGAM-CIFAR10 · CIFAR-100 test | PGD20 accuracy (%): 21.7 -> 21.7 | No clear effect |
| add adversarial gradient matching (lambda_adv=2 rather than 0) (ablation) | IGAM-CIFAR10 · CIFAR-100 test | PGD20 accuracy (%): 3.9 -> 21.7 | Helped |
| increase adversarial matching weight from 1 to 2 (ablation) | IGAM-CIFAR10 · CIFAR-100 test | PGD20 accuracy (%): 7.37 -> 21.7 | Helped |
| replace input resizing with center-cropping for teacher input adaptation | IGAM on Tiny-ImageNet · Tiny-ImageNet test | adversarial accuracy | No clear effect |
Closest papers by meaning
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Attribute-Guided Adversarial Training for Robustness to Natural Perturbations (AAAI 2021 · similarity 0.90)
- On Transfer of Adversarial Robustness from Pretraining to Downstream Tasks (NeurIPS 2023 · similarity 0.90)
- Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training (ICML 2024 · similarity 0.89)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.89)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · 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.