Research map / Adversarial robustness and security / Black-box and transfer attacks
Black-Box Dissector: Towards Erasing-Based Hard-Label Model Stealing Attack
Yixu Wang, Jie Li, Hong Liu, Yan Wang, Yongjian Wu, Feiyue Huang, Rongrong Ji
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 |
|---|---|---|---|
| add CAM-driven erasing (ablation) | black-box dissector · CIFAR10, 30k queries | test accuracy: 76.12% -> 80.47% | Improved |
| add self-KD module (ablation) | black-box dissector · CIFAR10, 30k queries | test accuracy: 79.02% -> 80.47% | Improved |
| use CAM-prior erasing instead of random erasing (CutOut) (ablation) | black-box dissector · CIFAR10 | test accuracy: 77.11% -> 80.47% | Improved |
| use self-KD instead of CutOut (ablation) | black-box dissector · CIFAR10 | test accuracy: 78.91% -> 80.47% | Improved |
| use self-KD instead of label smoothing (0.9) (ablation) | black-box dissector · CIFAR10 | test accuracy: 78.22% -> 80.47% | Improved |
| use self-KD instead of label smoothing (0.8) (ablation) | black-box dissector · CIFAR10 | test accuracy: 77.46% -> 80.47% | Improved |
| double weight decay on full method (ablation) | black-box dissector · CIFAR10 | test accuracy: 80.47% -> 81.65% | Mixed |
| add self-KD under adaptive misinformation (threshold 0.9) (ablation) | black-box dissector · CIFAR10 | test accuracy: 61.81% -> 74.40% | Improved |
| add self-KD under prediction poisoning (threshold 0.8) (ablation) | black-box dissector · CIFAR10 | test accuracy: 76.49% -> 79.23% | Improved |
| add k-Center selection on small-class datasets | black-box dissector · CIFAR10 | test accuracy: 80.47 -> 79.27 | Hurt |
| add k-Center selection on many-class dataset | black-box dissector · CUBS200 | test accuracy: 36.28 -> 44.43 | Improved |
| use black-box dissector instead of KnockoffNets | hard-label stealing · CIFAR10, transfer attack ResNet-34 | attack success rate: 57.85% -> 76.63% | Improved |
Closest papers by meaning
- Towards Data-Free Model Stealing in a Hard Label Setting (CVPR 2022 · similarity 0.89)
- Exploring Query Efficient Data Generation Towards Data-Free Model Stealing in Hard Label Setting (AAAI 2025 · similarity 0.86)
- Data-Free Hard-Label Robustness Stealing Attack (AAAI 2024 · similarity 0.86)
- DiffHammer: Rethinking the Robustness of Diffusion-Based Adversarial Purification (NeurIPS 2024 · similarity 0.85)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 2024 · similarity 0.85)
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.85)
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.
