Research map / Adversarial robustness and security / Black-box and transfer attacks
Defending against Data-Free Model Extraction by Distributionally Robust Defensive Training
Zhenyi Wang, Li Shen, Tongliang Liu, Tiehang Duan, Yanjun Zhu, Donglin Zhan, David Doermann, 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 |
|---|---|---|---|
| apply MeCo defense instead of no defense | DFME attack · CIFAR10, clone ResNet18-8X | clone accuracy: 87.36 ± 0.78% -> 51.68 ± 1.96% | Improved |
| apply MeCo instead of GRAD baseline | DFME attack · CIFAR100, clone ResNet18-8X | clone accuracy: 39.07 ± 2.72% -> 29.57 ± 1.97% | Improved |
| apply MeCo instead of no defense | MAZE attack · CIFAR10, clone ResNet18-8X | clone accuracy: 45.17 ± 0.73% -> 21.89 ± 2.07% | Improved |
| apply MeCo instead of no defense | DFMS-HL decision-based attack · CIFAR100, clone ResNet18-8X | clone accuracy: 72.57 ± 1.28% -> 59.30 ± 1.70% | Improved |
| output-perturbation baselines (RandP, P-poison, GRAD) instead of no defense | DFMS-HL decision-based attack · CIFAR10, clone ResNet18-8X | clone accuracy: 84.67 ± 1.90% -> 84.28 ± 1.95% | No clear effect |
| replace MeCo's small l1 perturbation with fixed l1 budget 1.0 (GRAD) | target model · CIFAR10 utility | l1 norm of probability perturbation: 1.0 -> 0.099 | Improved |
| MeCo input perturbation on benign accuracy | target model · CIFAR10 | test accuracy: 94.91 ± 0.37% -> 94.17 ± 0.56% | Hurt |
| attacker adapts by adding its own random perturbation generator | MeCo vs DFME · CIFAR10, clone ResNet18-8X | clone accuracy: 51.68 ± 1.96% -> 25.79 ± 0.81% | Improved |
| score-based attacker uses only hard label instead of probabilities | MeCo vs DFME · CIFAR10, clone ResNet18-8X | clone accuracy: 51.68 ± 1.96% -> 41.23 ± 0.58% | Improved |
| add DRO to defensive training (ablation) | MeCo · target model utility | model utility: 5.3% - 5.6% | Improved |
| increase gamma (ablation) | MeCo · CIFAR10 | utility vs defense | Mixed |
| increase number of test-data simulation steps Q (ablation) | MeCo · benign accuracy | benign accuracy | Hurt |
Closest papers by meaning
- Exploring Query Efficient Data Generation Towards Data-Free Model Stealing in Hard Label Setting (AAAI 2025 · similarity 0.87)
- Data-Free Hard-Label Robustness Stealing Attack (AAAI 2024 · similarity 0.86)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.86)
- HoneypotNet: Backdoor Attacks Against Model Extraction (AAAI 2025 · similarity 0.86)
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.86)
- From Counterfactuals to Trees: Competitive Analysis of Model Extraction Attacks (NeurIPS 2025 · similarity 0.86)
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.
