Research map / Adversarial robustness and security / Attacks on text and graphs
Bucks for Buckets (B4B): Active Defenses Against Stealing Encoders
Jan Dubiński, Stanisław Pawlak, Franziska Boenisch, Tomasz Trzcinski, Adam Dziedzic
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 coverage-adaptive noise and per-user transformations instead of returning unmodified representations | SimSiam ResNet50 API · CIFAR10, STL10, SVHN, and FashionMNIST legitimate-user tasks | accuracy: 90.41{ $\pm$0.02} -> 90.24{ $\pm$0.11}; 95.08{ $\pm$0.13} -> 95.05{ $\pm$0.1}; 75.47{ $\pm$0.04} -> 74.96{ $\pm$0.13}; 9 | No clear effect |
| Steal through B4B-protected API rather than an undefended API | SimSiam ResNet50 · CIFAR10, STL10, SVHN, and FashionMNIST; 50K ImageNet queries | accuracy: 65.2{ $\pm$0.03} -> 35.72{ $\pm$0.04}; 64.9{ $\pm$0.01} -> 31.54{ $\pm$0.02}; 63.1{ $\pm$0.01} -> 19.74{ $\pm$0.02}; 88. | Hurt |
| Increase stealing queries from 50K to 100K through the B4B-protected API | Single-account attacker · CIFAR10, STL10, SVHN, and FashionMNIST; ImageNet queries | accuracy: CIFAR10: 35.72{ $\pm$0.04} -> 12.01{ $\pm$0.07}; STL10: 31.54{ $\pm$0.02} -> 13.94{ $\pm$0.05}; SVHN: 19.74{ $\pm$0.02} | Mixed |
| Steal with LAION instead of ImageNet queries through B4B | SimSiam ResNet50 · CIFAR10 downstream task; 100K queries | accuracy: 64.92{ $\pm$0.03} -> 40.96{ $\pm$ 0.03} | Hurt |
| Use two sybil accounts and remap their representations | B4B-protected SimSiam stealing attack · ImageNet stealing; downstream classification tasks | accuracy: CIFAR10: 39.56{ $\pm$ 0.06}; STL10: 38.50{ $\pm$0.04}; SVHN: 23.41{ $\pm$0.02}; FashionMNIST: 77.01{ $\pm$ 0.08} | Mixed |
| Increase the number of sybil accounts from two to more accounts | B4B-protected stealing attack · ImageNet stealing; downstream classification tasks | accuracy: 2$\times$50k: 39.56{ $\pm$ 0.06}; 6$\times$16.7k: 26.62{ $\pm$ 0.04} | Hurt |
| Replace B4B with static noise at sigma 0.1 (ablation) | SimSiam API and 50K-query attacker · CIFAR10 | accuracy: Legitimate: 90.41 { $\pm$0.02} -> 90.20{ $\pm$0.03}; attacker: 65.2{ $\pm$0.03} -> 64.92{ $\pm$0.04} | No clear effect |
| Replace B4B with static noise at sigma 10 (ablation) | SimSiam API and 50K-query attacker · CIFAR10 | accuracy: Legitimate: 90.41 { $\pm$0.02} -> 65.11{ $\pm$0.45}; attacker: 65.2{ $\pm$0.03} -> 36.32{ $\pm$0.2} | Mixed |
| Use Binary rather than no transformation (ablation) | SimSiam representations for legitimate CIFAR10 users · CIFAR10 | accuracy: 90.41\% ($\pm$ 0.02) -> 88.78\% ($\pm$ 0.2) | Hurt |
| Increase lambda to 10^-4 from 10^-6 while keeping alpha=1 and beta=80% (ablation) | B4B cost function · Legitimate downstream tasks and ImageNet stealing | accuracy: Legitimate CIFAR10: 90.41{ $\pm$0.02} -> 90.02 { $\pm$0.1}; 50K attacker CIFAR10: 65.2 { $\pm$0.03} -> 28.22 { $\pm$ 0.0 | Mixed |
| Reduce beta from 50% to 30% (ablation) | B4B cost function · Legitimate CIFAR10 and 100K ImageNet stealing | accuracy: Legitimate CIFAR10: 90.27 { $\pm$0.07} -> 88.1 { $\pm$0.11}; attacker CIFAR10: 16.27 { $\pm$ 0.04} -> 9.57 { $\pm$ 0.17} | Mixed |
| Reduce alpha from 1 to 0.1 while keeping beta=30% and lambda=10^-6 (ablation) | B4B cost function · Legitimate CIFAR10 downstream task | accuracy: 88.1 { $\pm$0.11} -> 90.17 { $\pm$0.1} | Improved |
Closest papers by meaning
- Stealix: Model Stealing via Prompt Evolution (ICML 2025 · similarity 0.84)
- Hiding in Plain Sight: Disguising Data Stealing Attacks in Federated Learning (ICLR 2024 · similarity 0.83)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.82)
- The Good, the Bad and the Ugly: Meta-Analysis of Watermarks, Transferable Attacks and Adversarial Defenses (NeurIPS 2025 · similarity 0.82)
- GraphIP–Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It? (NeurIPS 2026 · similarity 0.82)
- Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game (ICLR 2024 · similarity 0.82)
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.
