Research map / Adversarial robustness and security / Black-box and transfer attacks
Deep Neural Network Fingerprinting by Conferrable Adversarial Examples
Nils Lukas, Yuxuan Zhang, Florian Kerschbaum
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 CEM-generated conferrable examples as the fingerprint instead of IPGuard examples | CIFAR-10 surrogate verification · CIFAR-10 retrained models | ROC AUC: 0.63 -> 1.0 | Improved |
| Generate examples with CEM at ε=0.01 rather than ε=0.15 | CEM · CIFAR-10 | mean conferrability: 0.24 -> 0.49 | Improved |
| Use a larger perturbation threshold, from ε=0.01 to ε=0.15 | CEM fingerprint · CIFAR-10 surrogate models | CAEAcc: 0.85 -> 0.98 | Improved |
| Use a larger perturbation threshold, from ε=0.01 to ε=0.15 | CEM fingerprint · CIFAR-10 reference models | CAEAcc: 0.42 -> 0.75 | Hurt |
| Detect fingerprint queries using the evaluated detector | CEM fingerprint verification · CIFAR-10, ε≤0.025 | ROC AUC: 0.67 | No clear effect |
| Apply fine-tuning and pruning to the source-derived model | CEM fingerprint · CIFAR-10 | fingerprint removability | No clear effect |
| Extract surrogates through retraining and related extraction attacks | CEM fingerprint · CIFAR-10 | fingerprint removability | No clear effect |
| Transfer learn from ImageNet32 using CIFAR-10 domain data | CEM fingerprint · CIFAR-10 | fingerprint removability | Hurt |
| Transfer learn without CIFAR-10 data, using CINIC data | CEM fingerprint · CIFAR-10 | fingerprint removability | No clear effect |
| Adversarially train the surrogate from scratch | CEM fingerprint · CIFAR-10, ε=0.025 | CAEAcc: 15% | Hurt |
| Give the extractor more ground-truth labels | CEM fingerprint · CIFAR-10 surrogate models | CAEAcc | Hurt |
| Set α=1.0 and β=γ=0.1 in the CEM objective (ablation) | CEM · CIFAR-10, ε=0.025 | mean conferrability: 0.40 -> 0.49 | Improved |
Closest papers by meaning
- Fingerprinting Deep Neural Networks Globally via Universal Adversarial Perturbations (CVPR 2022 · similarity 0.90)
- Sensitive-Sample Fingerprinting of Deep Neural Networks (CVPR 2019 · similarity 0.89)
- United We Stand, Divided We Fall: Fingerprinting Deep Neural Networks via Adversarial Trajectories (NeurIPS 2024 · similarity 0.87)
- SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models (CVPR 2025 · similarity 0.86)
- Queries, Representation & Detection: The Next 100 Model Fingerprinting Schemes (AAAI 2025 · similarity 0.85)
- IPRemover: A Generative Model Inversion Attack against Deep Neural Network Fingerprinting and Watermarking (AAAI 2024 · 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.
