Research map / Adversarial robustness and security / Certified and empirical robustness
Towards the first adversarially robust neural network model on MNIST
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 |
|---|---|---|---|
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L2 attacks | median adversarial distance (L2): 1.4 -> 2.3 | Helped |
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L2 attacks | accuracy at L2 epsilon = 1.5: 35 -> 80 | Helped |
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L∞ attacks | median adversarial distance (L∞): 0.34 -> 0.23 | Hurt |
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L∞ attacks | accuracy at L∞ epsilon = 0.3: 88 -> 8 | Hurt |
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L0 attacks | median adversarial distance (L0): 4.0 -> 22.0 | Helped |
| use class-conditional generative analysis-by-synthesis instead of adversarial training | MNIST classifier · MNIST test, all L0 attacks | accuracy at L0 epsilon = 12: 0 -> 78 | Helped |
| add test-time input binarization (ablation) | ABS · MNIST test, all L2 attacks | median adversarial distance (L2): 2.3 -> 1.3 | Hurt |
| add test-time input binarization (ablation) | ABS · MNIST test, all L∞ attacks | median adversarial distance (L∞): 0.23 -> 0.44 | Helped |
| add test-time input binarization (ablation) | ABS · MNIST test, all L0 attacks | median adversarial distance (L0): 22.0 -> 36.0 | Helped |
| add test-time input binarization (ablation) | ABS · MNIST test, L2 robustness guarantee | mean perturbation lower bound (L2): 0.690 ± 0.005 -> 0.601 ± 0.005 | Hurt |
| add test-time input binarization (ablation) | ABS · MNIST test, all L∞ attacks | accuracy at L∞ epsilon = 0.3: 8 -> 77 | Helped |
| add test-time input binarization (ablation) | ABS · MNIST test, all L0 attacks | accuracy at L0 epsilon = 12: 78 -> 82 | Helped |
Closest papers by meaning
- Evaluating Robustness of Neural Networks with Mixed Integer Programming (ICLR 2019 · similarity 0.91)
- Accurate, reliable and fast robustness evaluation (NeurIPS 2019 · similarity 0.91)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.91)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.91)
- Adversarial Training and Robustness for Multiple Perturbations (NeurIPS 2019 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
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.