Research map / Adversarial robustness and security / Certified and empirical robustness
Formalizing Generalization and Adversarial Robustness of Neural Networks to Weight Perturbations
Yu-Lin Tsai, Chia-Yi Hsu, Chia-Mu Yu, Pin-Yu Chen
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 |
|---|---|---|---|
| Add both the weight-perturbation robustness term and matrix-norm regularization to the classification loss | MNIST feed-forward network · MNIST, weight PGD attack (200 steps; varying ε_test) | robust-accuracy AUC: about 2× larger than that of the standard model | Improved |
| Add the weight-perturbation robustness term alone (ablation) | MNIST feed-forward network · MNIST, weight PGD attack (200 steps; varying ε_test) | robust accuracy | Improved |
| Add matrix-norm generalization-gap regularization alone (ablation) | MNIST feed-forward network · MNIST, weight PGD attack (200 steps; varying ε_test) | robust accuracy | Improved |
| Increase the matrix-norm regularization coefficient μ (ablation) | Proposed-loss-trained MNIST network · MNIST, empirical generalization gap under weight PGD attack | generalization gap | Improved |
| Replace the proposed matrix-norm regularization with L1 or L2 weight decay (ablation) | Proposed-loss-trained MNIST network · MNIST, empirical generalization gap and weight-perturbation robustness | generalization gap | Hurt |
| Use naive adversarial weight training instead of the proposed theory-driven loss | MNIST feed-forward network · MNIST, weight PGD attack at different ε_test | test accuracy | Hurt |
| Train with the proposed loss rather than the standard loss | MNIST network · MNIST, 2-bit weight quantization | accuracy decrease: 3x less than the standard model | Improved |
Closest papers by meaning
- Towards Certificated Model Robustness Against Weight Perturbations (AAAI 2020 · similarity 0.90)
- Sensitivity and Generalization in Neural Networks: an Empirical Study (ICLR 2018 · similarity 0.90)
- Adversarial Weight Perturbation Helps Robust Generalization (NeurIPS 2020 · similarity 0.89)
- The Price of Implicit Bias in Adversarially Robust Generalization (NeurIPS 2024 · similarity 0.88)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · similarity 0.88)
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks (NeurIPS 2025 · similarity 0.87)
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.
