Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Training from Mean Field Perspective
Soichiro Kumano, Hiroshi Kera, Toshihiko Yamasaki
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 |
|---|---|---|---|
| increase network width from 256 to 512 in adversarial training | vanilla ReLU network, L=20, eps=0.3 · MNIST training accuracy | adversarial trainability: N = 256 -> 512 | Improved |
| add shortcuts (residual connections) to the network | deep fully connected ReLU-like network · adversarial training trainability | gradient trainability | Improved |
| use adversarial training instead of standard training | vanilla network L=10, p=inf, q=inf, eps=0.3 · MNIST | weight variance | Hurt |
| increase width N from 1k to 2k | adversarial training of vanilla network · MNIST weight variance | weight variance decrease: N = 1k -> N = 2k | Improved |
| increase depth L under adversarial training | vanilla network · MNIST training accuracy | training accuracy | Hurt |
| use standard training instead of adversarial training | vanilla networks of various depth and width · MNIST training accuracy | training accuracy | Improved |
| increase network width in adversarial training | fully trained networks · MNIST, Fashion-MNIST | adversarial robustness | Improved |
| increase training steps T from 100 to 160 in trainability boundary | vanilla networks p=inf q=inf eps=0.3 · MNIST depth-width heat map | trainability boundary: T = 100, 160, 250 | Mixed |
| increase input/output dimension for (p,q)=(2,inf) | adversarial loss upper bound · random networks | upper bound | No clear effect |
| use finite network width instead of infinite | Thm 5.1 bounds · random networks N=40,000 | adversarial loss vs bound | Mixed |
Closest papers by meaning
- Over-parameterized Adversarial Training: An Analysis Overcoming the Curse of Dimensionality (NeurIPS 2020 · similarity 0.89)
- Adversarial Robustness Guarantees for Random Deep Neural Networks (ICML 2021 · similarity 0.89)
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (NeurIPS 2024 · similarity 0.87)
- A Closer Look at Curriculum Adversarial Training: From an Online Perspective (AAAI 2024 · 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.
