Research map / Label-efficient and robust learning / Recognition and long-tail
Dual Student Networks for Data-Free Model Stealing
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 |
|---|---|---|---|
| Dual Students generator loss (two students, disagreement objective) replacing DFME forward-differences generator loss | data-free model stealing, soft-label setting · MNIST, FashionMNIST, GTSRB, SVHN, CIFAR10, CIFAR100 | student accuracy (probabilities): e.g. FashionMNIST 85.17 -> 91.17; CIFAR100 26.46 -> 45.32 | Helped |
| Dual Students generator loss replacing DFME generator loss | data-free model stealing, hard-label setting · FashionMNIST, CIFAR10, CIFAR100 | student accuracy (hard-labels): FashionMNIST 48.91 -> 80.53; CIFAR10 68.40 -> 78.72; CIFAR100 6.91 -> 9.77 | Helped |
| Dual Students fine-tuning of fully trained DFMS-SL/HL student (ablation) | DFMS-SL/HL · CIFAR10 and CIFAR100 | student accuracy (hard-labels): CIFAR10 79.61 -> 85.06; CIFAR100 35.78 -> 36.38 | Helped |
| Dual Students fine-tuning of DFMS-SL/HL (ablation) | DFMS-SL/HL · CIFAR10 and CIFAR100 | student accuracy (probabilities): CIFAR10 88.51 -> 89.38; CIFAR100 44.86 -> 50.98 | Helped |
| Dual Students generator loss replacing DFME generator loss | transfer-based PGD attack proxy (soft-label student) · CIFAR10, eps=3/255, untargeted | attack success %: 83.59 -> 91.04 | Helped |
| Dual Students generator loss replacing DFME generator loss | transfer-based FGSM attack proxy (soft-label student) · CIFAR10, eps=3/255, untargeted | attack success %: 56.84 -> 62.35 | Helped |
| Dual Students fine-tuning of DFMS-HL/SL student (ablation) | DFMS-HL/SL proxy for PGD attack (hard-label) · CIFAR10, eps=3/255, untargeted | attack success %: 72.40 -> 73.08 | Helped |
| Dual Students generator loss on CIFAR10 soft-label setting with more balanced generated class distribution | generator class balance · CIFAR10 | class distribution of generated samples | Helped |
| Dual Students with class count unknown (classes discovered as generator explores, expanding classifier) (ablation) | data-free hard-label setting without known number of classes · CIFAR10 | student accuracy (hard-labels): 74.95 vs ~78.72 with known classes (comparable) | No clear effect |
| Dual Students generator loss on hard-label setting for FashionMNIST and GTSRB | data-free hard-label student · GTSRB, SVHN | student accuracy (hard-labels): GTSRB 85.69 -> 93.20; SVHN 93.87 -> 95.43 | Helped |
| Dual Students generator loss, fewer target queries (no zeroth-order gradient estimation) | query efficiency · CIFAR10 soft-label | queries relative to DFME: 5/7 (~71.42%) of DFME queries | Helped |
| Dual Students generator loss on SVHN soft-label setting | data-free student · SVHN | student accuracy (probabilities): 95.33 -> 95.72 | Helped |
Closest papers by meaning
- Exploring Query Efficient Data Generation Towards Data-Free Model Stealing in Hard Label Setting (AAAI 2025 · similarity 0.90)
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.89)
- Fully Exploiting Every Real Sample: SuperPixel Sample Gradient Model Stealing (CVPR 2024 · similarity 0.88)
- Efficient Model Stealing Defense with Noise Transition Matrix (CVPR 2024 · similarity 0.88)
- SNEAKDOOR: Stealthy Backdoor Attacks against Distribution Matching-based Dataset Condensation (NeurIPS 2025 · similarity 0.85)
- Adversarial Machine Unlearning (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.