Research map / Label-efficient and robust learning / Recognition and long-tail
Defending against Model Stealing via Verifying Embedded External Features
Yiming Li, Linghui Zhu, Xiaojun Jia, Yong Jiang, Shu-Tao Xia, Xiaochun Cao
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 |
|---|---|---|---|
| Use style transfer to embed external features instead of a white-square patch (ablation) | Ownership verification against model stealing · CIFAR-10 / Distillation, Zero-shot, Fine-tuning, Label-query, Logit-query | p-value: 0.17 -> 10^{-7}; 0.01 -> 10^{-5}; 10^{-3} -> 10^{-6}; 10^{-3} -> 10^{-4}; 10^{-3} -> 10^{-4} | Improved |
| Add the meta-classifier to BadNets verification (ablation) | BadNets-based model watermarking · CIFAR-10 / Distillation, Zero-shot, Fine-tuning, Label-query, Logit-query | p-value: 0.32 -> 10^{-3}; 0.22 -> 10^{-61}; 0.28 -> 10^{-5}; 0.20 -> 10^{-50}; 0.23 -> 10^{-3} | Improved |
| Use the sign of gradients rather than raw gradients as meta-classifier input (ablation) | Ownership meta-classifier · CIFAR-10 and ImageNet / model-stealing attacks | p-value: CIFAR-10: 0.01 -> 10^{-7}; ImageNet: 10^{-4} -> 10^{-5} | Improved |
| Increase the transformation rate (ablation) | Victim model training and ownership verification · Model-stealing attacks | p-value | Improved |
| Increase the number of sampled transformed images (ablation) | Hypothesis-test verification · Ownership verification | p-value | Improved |
| Change the style image (ablation) | External-feature embedding · CIFAR-10 / model-stealing attacks | p-value | No clear effect |
| Use testing images instead of training images to generate verification samples (ablation) | Ownership verification · CIFAR-10 and ImageNet / model-stealing attacks | confidence score (Delta mu) | No clear effect |
| Use style-transformed samples to train the victim model | Victim model · CIFAR-10 / accuracy | accuracy: 91.99\% -> 91.79\% | No clear effect |
| Use BadNets rather than external-feature verification | Model stealing defense · CIFAR-10 / stolen model after zero-shot distillation | attack success rate: 98.02 -> 3.84 | Hurt |
| Apply dataset inference to unrelated models trained on disjoint subsets | Dataset inference · CIFAR-10 / ownership verification | p-value: 10^{-7}; 10^{-5}; 10^{-4} | Hurt |
| Use external-feature verification rather than the listed model-watermarking baselines | Ownership verification · CIFAR-10 and ImageNet / model-stealing attacks | p-value: CIFAR-10 distillation: 0.33 -> 10^{-7}; ImageNet distillation: 0.19 -> 10^{-5} | Improved |
Closest papers by meaning
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.88)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.86)
- Efficient Model Stealing Defense with Noise Transition Matrix (CVPR 2024 · similarity 0.86)
- Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy (AAAI 2025 · similarity 0.86)
- Fully Exploiting Every Real Sample: SuperPixel Sample Gradient Model Stealing (CVPR 2024 · similarity 0.85)
- SAME: Sample Reconstruction against Model Extraction Attacks (AAAI 2024 · 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.
