Research map / Adversarial robustness and security / Black-box and transfer attacks
Training A Secure Model against Data-Free Model Extraction
Zhenyi Wang, Li Shen, junfeng guo, Tiehang Duan, Siyu Luan, Tongliang Liu, Mingchen Gao
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 |
|---|---|---|---|
| train victim with AAUG (weight perturbation plus attack-aware training) instead of output-perturbation defenses | ResNet34-8x victim · CIFAR-10 DFME soft-label | clone model accuracy (ResNet18-8x): 87.36 -> 46.82 | Improved |
| train victim with AAUG | ResNet34-8x victim · CIFAR-100 DFME soft-label | clone model accuracy (ResNet18-8x): 58.72 -> 21.31 | Improved |
| train victim with AAUG | ResNet34-8x victim · ImageNet-100 DFME soft-label | clone model accuracy (ResNet18-8x): 35.89 -> 15.58 | Improved |
| apply output-perturbation defenses (RandP, P-poison, GRAD) against hard-label DFMS-HL | ResNet34-8x victim · CIFAR-10 DFMS-HL | clone model accuracy (ResNet18-8x): 84.67 -> 84.28 (GRAD) | No clear effect |
| train victim with AAUG against hard-label DFMS-HL | ResNet34-8x victim · CIFAR-100 DFMS-HL | clone model accuracy (ResNet18-8x): 72.57 -> 45.68 | Improved |
| train victim with AAUG | ResNet34-8x victim · CIFAR-10 victim utility | test accuracy: 94.91 -> 94.21 | No clear effect |
| train victim with AAUG instead of output perturbation | ResNet34-8x victim · CIFAR-100 victim utility | test accuracy: 76.71 -> 75.73 | No clear effect |
| use AAUG instead of P-poison and GRAD at test time | defense deployment · CIFAR-10 and CIFAR-100 | test-time speed: 25 ~103x speed up | Improved |
| use AAUG instead of standard base model training | victim model training · training cost | training cost: 1.2x | Hurt |
| fine-tune pre-trained model with AAUG for 20 epochs instead of 200-epoch retraining (ablation) | AAUG · CIFAR-10 | defense performance and cost: training cost reduced by 83% | No clear effect |
| increase attacker query budget (ablation) | baseline defenses vs AAUG · DFME attack | clone model accuracy | Mixed |
Closest papers by meaning
- Exploring Query Efficient Data Generation Towards Data-Free Model Stealing in Hard Label Setting (AAAI 2025 · similarity 0.88)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.87)
- HoneypotNet: Backdoor Attacks Against Model Extraction (AAAI 2025 · similarity 0.86)
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.86)
- SAME: Sample Reconstruction against Model Extraction Attacks (AAAI 2024 · similarity 0.86)
- From Counterfactuals to Trees: Competitive Analysis of Model Extraction Attacks (NeurIPS 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.
