Research map / Adversarial robustness and security / Black-box and transfer attacks
Subspace Attack: Exploiting Promising Subspaces for Query-Efficient Black-box Attacks
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 reference-model input gradients to span promising search subspaces | black-box Subspace Attack · CIFAR-10 and ImageNet, untargeted | mean queries: 713 -> 392; 373 -> 250; 1160 -> 555; 887 -> 462; 1437 -> 680; 1055 -> 456 | Helped |
| use reference-model input gradients to span promising search subspaces | black-box Subspace Attack · CIFAR-10 and ImageNet, untargeted | failure rate: 1.2% -> 0.3%; 0.0% -> 0.0%; 1.2% -> 0.7%; 4.2% -> 1.1%; 12.1% -> 4.2%; 6.4% -> 1.9% | Helped |
| increase the maximum dropout ratio from 0.0 to 0.5 (ablation) | Subspace Attack with CIFAR-10.1-trained reference models · CIFAR-10, WRN | failure rate: 9.6% -> 0.3% | Helped |
| increase the maximum dropout ratio from 0.0 to 0.2 (ablation) | Subspace Attack with CIFAR-10.1-trained reference models · CIFAR-10, WRN | mean queries: 519 -> 380 | Helped |
| increase reference training data from CIFAR-10.1 (2k images) to CIFAR-10.1 plus 8k CIFAR-10 test images (ablation) | Subspace Attack, no dropout · CIFAR-10, WRN | mean queries: 519 -> 239 | Helped |
| increase reference training data from CIFAR-10.1 (2k images) to CIFAR-10.1 plus 8k CIFAR-10 test images (ablation) | Subspace Attack, no dropout · CIFAR-10, WRN | failure rate: 9.6% -> 3.2% | Helped |
| add more reference models, from VGG-19 to VGG-19/16/13 (ablation) | Subspace Attack · CIFAR-10, WRN | mean queries: 400 -> 395 | Helped |
| add more reference models, from VGG-19/16/13 to VGG-19/16/13/11+AlexNet (ablation) | Subspace Attack · CIFAR-10, WRN | failure rate: 0.4% -> 0.3% | Helped |
| compute all five prior gradients and search the complete subspace instead of coordinate descent (ablation) | Subspace Attack · CIFAR-10, WRN | mean queries: 392 -> 389 | No clear effect |
Closest papers by meaning
- Query-Efficient and Scalable Black-Box Adversarial Attacks on Discrete Sequential Data via Bayesian Optimization (ICML 2022 · similarity 0.89)
- Prior Convictions: Black-box Adversarial Attacks with Bandits and Priors (ICLR 2019 · similarity 0.89)
- Simple Black-box Adversarial Attacks (ICML 2019 · similarity 0.89)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.88)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 2024 · similarity 0.88)
- Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided by a Function Prior (ICML 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.