Research map / Label-efficient and robust learning / Self-supervised and contrastive
Dataset Inference for Self-Supervised Models
Adam Dziedzic, Haonan Duan, Muhammad Ahmad Kaleem, Nikita Dhawan, Jonas Guan, Yannis Cattan, Franziska Boenisch, Nicolas Papernot
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 |
|---|---|---|---|
| increase stealing queries from 5K to 250K (ImageNet queries) (ablation) | dataset inference · ImageNet victim | p-value: 9.88e-1 -> 3.14e-5 | Improved |
| increase stealing queries from 500 to 50K (CIFAR10 queries) (ablation) | stolen encoder, SVHN victim · SVHN victim | p-value: 6.89e-1 -> 1.19e-17 | Improved |
| steal with the victim's own data type (SVHN queries on SVHN victim) vs other data | dataset inference · SVHN victim | p-value: 1.19e-17 -> 1.05e-75 | Improved |
| apply shuffle obfuscation to stolen encoder (ablation) | dataset inference · CIFAR10 victim | p-value: 8.73e-7 -> 1.72e-6 | No clear effect |
| apply pad obfuscation to stolen encoder (ablation) | dataset inference · CIFAR10 victim | p-value: 8.73e-7 -> 3.44e-6 | No clear effect |
| apply linear transform obfuscation (ablation) | dataset inference · CIFAR10 victim | p-value: 8.73e-7 -> 6.81e-7 | No clear effect |
| shuffle the representation elements (ablation) | cosine similarity score · CIFAR10 victim | cosine similarity C: 0.504 -> 0.002 | Hurt |
| pad the representation with zeros (ablation) | cosine similarity score · CIFAR10 victim | cosine similarity C: 0.504 -> 0.007 | Hurt |
| shuffle the representation elements (ablation) | mutual information score · CIFAR10 victim | mutual information score S: 0.73 -> 0.74 | No clear effect |
| use independent encoder instead of stolen (cosine similarity) | encoder similarity · CIFAR10 victim | cosine similarity C: 0.95 -> 0.0007 | Hurt |
| use independent encoder instead of stolen (mutual information) | encoder similarity · SVHN victim | mutual information score S: 0.96 -> 0.90 | No clear effect |
| increase queries from 500 to 50K (CIFAR10 queries) (ablation) | stolen encoder · SVHN downstream | linear eval accuracy: 19.3 -> 75.1 | Improved |
Closest papers by meaning
- Dataset Inference: Ownership Resolution in Machine Learning (ICLR 2021 · similarity 0.93)
- On the Difficulty of Defending Self-Supervised Learning against Model Extraction (ICML 2022 · similarity 0.91)
- Towards Data-Free Model Stealing in a Hard Label Setting (CVPR 2022 · similarity 0.88)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.87)
- Data Taggants: Dataset Ownership Verification Via Harmless Targeted Data Poisoning (ICLR 2025 · similarity 0.86)
- Stealix: Model Stealing via Prompt Evolution (ICML 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.
