Research map / Adversarial robustness and security / Black-box and transfer attacks
Adversarial Machine Learning at Scale
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 |
|---|---|---|---|
| Adversarially train Inception v3 with step l.l. examples instead of standard training | Inception v3 · ImageNet validation, ε=16 step l.l. | top-1 accuracy: 29.5% -> 73.9% | Helped |
| Adversarially train Inception v3 with step l.l. examples instead of standard training | Inception v3 · ImageNet validation, clean | top-1 accuracy: 78.4% -> 77.6% | Hurt |
| Add two Inception blocks to the adversarially trained model | Inception v3 with step l.l. adversarial training · ImageNet validation, ε=16 step l.l. | top-1 accuracy: 73.9% -> 74.4% | Helped |
| Train with one-step step l.l. examples rather than standard training | Inception v3 · ImageNet validation, ε=16 iterative least-likely | top-1 accuracy: 0.7% -> 1.5% | Helped |
| Train with one-step step l.l. examples rather than standard training | Inception v3 · ImageNet validation, ε=16 basic iterative | top-1 accuracy: 25.9% -> 23.2% | Hurt |
| Use FGSM rather than step l.l. for adversarial training and evaluation (ablation) | Inception v3 · ImageNet validation, ε=16 FGSM | top-1 accuracy: 72.0% -> 87.0% | Mixed |
| Increase the number of adversarial examples in the minibatch from k=16 to k=32 (ablation) | Inception v3 adversarial training · ImageNet validation, ε=16 step l.l. | top-1 accuracy: 75.4% -> 75.8% | No clear effect |
| Increase the number of adversarial examples in the minibatch from k=16 to k=32 (ablation) | Inception v3 adversarial training · ImageNet validation, clean | top-1 accuracy: 77.6% -> 76.3% | Hurt |
| Replace relu with relu6 (ablation) | Inception v3 adversarial training · ImageNet validation, ε=16 basic iterative | top-1 accuracy: 21.0% -> 23.2% | Helped |
| Replace relu with tanh (ablation) | Inception v3 adversarial training · ImageNet validation, ε=16 step l.l. | top-1 accuracy: 74.5% -> 61.9% | Hurt |
| Delay adversarial training by 10k iterations (ablation) | Inception v3 · ImageNet validation, clean | accuracy | No clear effect |
| Delay adversarial training by larger amounts (ablation) | Inception v3 · ImageNet validation, high-magnitude adversarial examples | accuracy: up to 4% | Hurt |
Closest papers by meaning
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.92)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.90)
- Ensemble Adversarial Training: Attacks and Defenses (ICLR 2018 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- A Direct Approach to Robust Deep Learning Using Adversarial Networks (ICLR 2019 · similarity 0.90)
- Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · 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.