Research map / Adversarial robustness and security / Black-box and transfer attacks
Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models
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 the decision-only Boundary Attack instead of gradient-based attacks | MNIST untargeted · MNIST | median squared L2-distance: 2.2e-03 (Carlini & Wagner) vs 3.6e-03 (Boundary); 4.3e-03 (DeepFool); 4.2e-02 (FGSM) | Mixed |
| Use the decision-only Boundary Attack instead of gradient-based attacks | CIFAR-10 untargeted · CIFAR-10 | median squared L2-distance: 5.6e-06 (Boundary) vs 5.8e-06 (DeepFool), 7.5e-06 (Carlini & Wagner), 2.5e-05 (FGSM) | Mixed |
| Use the decision-only Boundary Attack instead of gradient-based attacks | ImageNet-1000 untargeted · ImageNet-1000 | median squared L2-distance: VGG-19: 2.9e-07 vs 1.9e-07 (DeepFool), 5.7e-07 (Carlini & Wagner), 1.0e-06 (FGSM); ResNet-50: 1.0e-07 vs 7.5e-08, 2.2e-0 | Mixed |
| Use the decision-only Boundary Attack instead of Carlini & Wagner | Targeted attack · MNIST | median squared L2-distance: 4.8e-03 -> 6.5e-03 | Hurt |
| Use the decision-only Boundary Attack instead of Carlini & Wagner | Targeted attack · CIFAR-10 | median squared L2-distance: 3.0e-05 -> 3.3e-05 | Hurt |
| Use the decision-only Boundary Attack instead of Carlini & Wagner | Targeted attack · ImageNet-1000 | median squared L2-distance: 5.7e-06 -> 9.9e-06 | Hurt |
| Apply defensive distillation | Boundary Attack on MNIST · MNIST | median squared L2-distance: 3.6e-03 -> 4.2e-03 | No clear effect |
| Apply defensive distillation | Boundary Attack on CIFAR-10 · CIFAR-10 | median squared L2-distance: 5.6e-06 -> 1.3e-05 | No clear effect |
| Apply defensive distillation | FGSM · MNIST and CIFAR-10 | attack success: 4.2e-02 -> fails; 2.5e-05 -> fails | Hurt |
Closest papers by meaning
- ADBA: Approximation Decision Boundary Approach for Black-Box Adversarial Attacks (AAAI 2025 · similarity 0.90)
- Black-box Adversarial Attacks with Limited Queries and Information (ICML 2018 · similarity 0.90)
- QEBA: Query-Efficient Boundary-Based Blackbox Attack (CVPR 2020 · similarity 0.89)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.89)
- Mind the Gap: Detecting Black-box Adversarial Attacks in the Making through Query Update Analysis (CVPR 2025 · similarity 0.88)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 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.