Research map / Federated learning and privacy / Privacy attacks and secure learning
One Trap to Block Them All: Defending Encoder Stealing via Isotropic Uniformity
YITONG SHI, Kang Wei, Fushuo Huo, Shuchi Wu, Chuan Ma
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 UniTrap defense (eps 0.3) against RDA | SSL encoder · CIFAR-10 | RDA stolen accuracy (%): 77.90 -> 35.26 | Improved |
| apply UniTrap defense against RDA | SSL encoder · SVHN | RDA stolen accuracy (%): 73.85 -> 25.96 | Improved |
| apply UniTrap with ImageNet target encoder | SSL encoder · SVHN (ImageNet target) | RDA stolen accuracy (%): 60.03 -> 19.81 | Improved |
| apply UniTrap defense | SSL encoder · CIFAR-10 | target accuracy (%): 91.20 -> 90.18 | No clear effect |
| replace cosine distance-oriented perturbation with Gaussian-potential uniformity (ablation) | UniTrap · CIFAR-10 | StolenEncoder stolen accuracy (%): 44.10 -> 36.27 | Improved |
| replace cosine distance-oriented perturbation with uniformity (ablation) | UniTrap · STL-10 | StolenEncoder stolen accuracy (%): 47.67 -> 37.79 | Improved |
| replace cosine distance-oriented perturbation with uniformity (ablation) | UniTrap · CIFAR-10 | RDA stolen accuracy (%): 40.35 -> 35.26 | Improved |
| replace cosine distance-oriented perturbation with uniformity (ablation) | UniTrap · F-MNIST | RDA stolen accuracy (%): 79.79 -> 81.20 | No clear effect |
| replace cosine distance-oriented perturbation with uniformity (ablation) | UniTrap · CIFAR-10 | target accuracy (%): 90.19 -> 90.18 | No clear effect |
| replace cosine with uniformity against ContSteal (ablation) | UniTrap · CIFAR-10 | ContSteal stolen accuracy (%): 37.38 -> 37.77 | No clear effect |
| UniTrap vs Top-k defense at similar stolen accuracy | UniTrap · CIFAR-10 | target accuracy (%): 73.44 -> 90.18 | Improved |
| add Gaussian noise 0.5 as defense instead of UniTrap | Gaussian noise · CIFAR-10 | target accuracy (%): 90.18 -> 71.85 | Hurt |
Closest papers by meaning
- Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders (CVPR 2023 · similarity 0.89)
- Differentially private optimization for non-decomposable objective functions (ICLR 2025 · similarity 0.86)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.85)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.84)
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.84)
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.84)
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.
