Research map / Adversarial robustness and security / Attacks on text and graphs
Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders
Zeyang Sha, Xinlei He, Ning Yu, Michael Backes, Yang Zhang
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 contrastive stealing rather than conventional embedding-matching | MoCo ResNet18 encoder stealing · CIFAR10 downstream; STL10 surrogate | agreement: 0.479 -> 0.841 | Improved |
| Use contrastive stealing rather than conventional embedding-matching | SimCLR ResNet18 encoder stealing · STL10 downstream; STL10 surrogate | accuracy: 0.457 -> 0.714 | Improved |
| Use contrastive stealing rather than INFONCE | Image encoder stealing · CIFAR10 downstream; CIFAR10 surrogate | agreement: 0.826 -> 0.845 | Improved |
| Use contrastive stealing rather than INFONCE | Image encoder stealing · STL10 downstream; STL10 surrogate | agreement: 0.806 -> 0.829 | Improved |
| Replace training from scratch with Cont-Steal | SimCLR ResNet18 encoder · CIFAR10 pretraining | training time (h): 20.01 -> 0.62 | Improved |
| Add Gaussian noise with standard deviation 10 to target embeddings | Cont-Steal · STL10 surrogate and downstream; SimCLR pretrained on CIFAR10 | accuracy: 0.729 -> 0.410 | Hurt |
| Add negative pairs generated by the surrogate encoder (ablation) | Cont-Steal with BYOL · STL10 downstream; STL10 surrogate | agreement: 0.314 -> 0.817 | Improved |
| Add target-encoder negative pairs (ablation) | Cont-Steal with BYOL · CIFAR10 surrogate | agreement: 0.242 -> 0.844 | Improved |
| Add target-encoder negative pairs (ablation) | Cont-Steal with BYOL · STL10 surrogate | agreement: 0.314 -> 0.817 | Improved |
| Use a ResNet50 surrogate instead of ResNet18 (ablation) | Cont-Steal on SimCLR · CIFAR10 surrogate | accuracy: 0.839 -> 0.840 | No clear effect |
| Use DenseNet161 rather than ResNet18 as the surrogate (ablation) | Cont-Steal on SimCLR · CIFAR10 surrogate | accuracy: 0.839 -> 0.828 | Hurt |
| Use MobileNetV2 rather than ResNet18 as the surrogate (ablation) | Cont-Steal on SimCLR · CIFAR10 surrogate | accuracy: 0.839 -> 0.811 | Hurt |
Closest papers by meaning
- One Trap to Block Them All: Defending Encoder Stealing via Isotropic Uniformity (ECCV 2026 · similarity 0.89)
- Defending against Model Stealing via Verifying Embedded External Features (AAAI 2022 · similarity 0.87)
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.84)
- Data Poisoning based Backdoor Attacks to Contrastive Learning (CVPR 2024 · similarity 0.84)
- Reinforced Multi-teacher Knowledge Distillation for Efficient General Image Forgery Detection and Localization (AAAI 2025 · similarity 0.83)
- CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification (ICLR 2025 · similarity 0.83)
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.
