Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness is at Odds with Lazy Training
Yunjuan Wang, Enayat Ullah, Poorya Mianjy, Raman Arora
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 |
|---|---|---|---|
| Constrain trained weights to remain in the lazy regime | Two-layer ReLU networks · Any data distribution on the unit sphere | robust error: L_R(\textrm{W};\mathcal{D})\geq 0.9 | Hurt |
| Increase the allowed deviation of weights from initialization (ablation) | SGD-trained MNIST networks · MNIST 0-vs-1, varying input dimension | minimum perturbation size needed to flip prediction | Mixed |
| Increase the perturbation budget beyond 5/\sqrt{d} (ablation) | Adversarially trained lazy-regime network · MNIST 0-vs-1, varying input dimension | robust accuracy: 5/\sqrt{d} -> 0 | Hurt |
| Allow the weight radius to exceed the lazy-regime scale (ablation) | Adversarially trained network · MNIST 0-vs-1, d = 28^2 | robust test accuracy: 25/\sqrt{m} | Improved |
| Keep the weight radius below the phase-transition scale (ablation) | Adversarially trained network · MNIST 0-vs-1, d = 28^2 | robust test accuracy: 0 | Hurt |
| Project adversarial-training updates to keep weights in the lazy regime (ablation) | Adversarially trained MNIST networks · MNIST 0-vs-1 | test accuracy | No clear effect |
| Train beyond the lazy regime (ablation) | Adversarially trained network · MNIST 0-vs-1 | robust test accuracy | Improved |
Closest papers by meaning
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.91)
- Scaling up the Randomized Gradient-Free Adversarial Attack Reveals Overestimation of Robustness Using Established Attacks (IJCV 2019 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- Accurate, reliable and fast robustness evaluation (NeurIPS 2019 · similarity 0.89)
- Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients (AAAI 2018 · similarity 0.89)
- Adversarial Training Should Be Cast as a Non-Zero-Sum Game (ICLR 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.
