Research map / Adversarial robustness and security / Black-box and transfer attacks
Practical No-box Adversarial Attacks against DNNs
Qizhang Li, Yiwen Guo, Hao 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 |
|---|---|---|---|
| Replace conventional small-data supervised training with prototypical reconstruction | No-box substitute models · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 81.29% -> 33.94% | Improved |
| Add multiple decoders to prototypical reconstruction (ablation) | Prototypical no-box substitute · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 33.94% -> 31.26% | Improved |
| Train with rotation reconstruction rather than conventional unsupervised autoencoding | No-box substitute models · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 57.78% -> 45.35% | Improved |
| Train with jigsaw reconstruction rather than conventional unsupervised autoencoding | No-box substitute models · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 57.78% -> 46.04% | Improved |
| Increase prototypical decoder count from 1 to 5 (ablation) | Prototypical no-box substitute · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 33.94% -> 32.05% | Improved |
| Increase prototypical decoder count from 5 to 10 (ablation) | Prototypical no-box substitute · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 32.05% -> 31.50% | Improved |
| Increase prototypical decoder count from 10 to 20 (ablation) | Prototypical no-box substitute · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 31.50% -> 31.26% | Improved |
| Use prototypical reconstruction rather than Beyonder transfer | No-box attacks · ImageNet, ε=0.1, average across victim models | prediction accuracy on adversarial examples: 36.26% -> 31.26% | Improved |
| Use the no-box attack instead of Beyonder transfer | Clarifai celebrity recognition · Clarifai, ε=0.1 | prediction accuracy on adversarial examples: 14.29% -> 15.40% | Hurt |
| Use the no-box attack instead of Beyonder transfer | Clarifai celebrity recognition · Clarifai, ε=0.08 | prediction accuracy on adversarial examples: 21.85% -> 33.67% | Hurt |
| Apply proposed no-box adversarial examples rather than Naïve† examples | Adversarially trained ResNet · ImageNet | accuracy on adversarial examples: 54.12% -> 39.24% | Improved |
| Increase the number of training images (ablation) | Prototypical no-box substitutes · ImageNet | prediction accuracy on adversarial examples | Improved |
Closest papers by meaning
- Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains (ICLR 2022 · similarity 0.89)
- NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks (ICML 2019 · similarity 0.88)
- Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks (CVPR 2019 · similarity 0.88)
- Improving Transferability of Adversarial Examples With Input Diversity (CVPR 2019 · similarity 0.88)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.88)
- Data-Free Universal Attack by Exploiting the Intrinsic Vulnerability of Deep Models (AAAI 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.
