Research map / Adversarial robustness and security / Certified and empirical robustness
First-Order Adversarial Vulnerability of Neural Networks and Input Dimension
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 |
|---|---|---|---|
| Increase input dimension | Standard feedforward networks after training · Upsampled CIFAR-10 test set | adversarial vulnerability | Hurt |
| Increase PGD regularization | CNNs · Upsampled CIFAR-10 test set | dimension dependence of adversarial vulnerability | Helped |
| Increase PGD regularization | CNNs · CIFAR-10 | accuracy | No clear effect |
| Use PGD training rather than down-sampling | CNNs · Mini-ImageNet | accuracy-vulnerability trade-off | Helped |
| Use average-pooling instead of strided or max-pooling layers (ablation) | CNNs · CIFAR-10 | adversarial robustness | Helped |
| Replace strided convolutions with non-strided convolutions followed by average-pooling (ablation) | CNNs · CIFAR-10 | input-to-output gradient magnitude | Helped |
| Use gradient regularization instead of adversarial augmentation, with the corresponding dual norm (ablation) | CNNs · CIFAR-10 | adversarial vulnerability | No clear effect |
| Use ℓ1 rather than ℓ2 loss-gradient regularization (ablation) | CNNs · CIFAR-10 | accuracy-vulnerability trade-off | No clear effect |
| Use stronger cross-Lipschitz regularization rather than gradient regularization or adversarial augmentation (ablation) | CNNs · CIFAR-10 | accuracy-vulnerability trade-off | Hurt |
| Continue training after training loss reaches its minimum | Neural networks · Test set | adversarial robustness | Hurt |
| Apply early stopping | Neural networks · Test set | adversarial vulnerability | Helped |
| Increase regularization beyond the point where dimension dependence is nearly broken (ablation) | CNNs · CIFAR-10 | accuracy | Hurt |
Closest papers by meaning
- With Friends Like These, Who Needs Adversaries? (NeurIPS 2018 · similarity 0.92)
- Excessive Invariance Causes Adversarial Vulnerability (ICLR 2019 · similarity 0.90)
- Adversarial Vulnerability from Interference Between Features in Superposition (ICML 2026 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.88)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · 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.