Research map / Label-efficient and robust learning / Recognition and long-tail
Fingerprinting Pre-trained Encoders under Arbitrary Downstream Fine-Tuning via Adversarial Shifting
Tianlong Xu, Zixiong Wang, Lishuai Hou, Gaoyang Liu, Chen Wang, Xiaoyi Fan
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 |
|---|---|---|---|
| switch downstream adaptation from head-only (FTLL) to full fine-tuning (FTAL) | Adversarial Shifting · SimCLR ResNet-50 -> STL-10 | Matching Rate: 0.960 -> 0.760 | Hurt |
| switch from FTLL to FTAL | SSL-WM · CIFAR-100 | Matching Rate: 0.484 -> 0.291 | Hurt |
| use label-only group fingerprint instead of label-bound fingerprints | Adversarial Shifting vs ADV-TRA/MFUE · FTLL, SimCLR STL-10 | Matching Rate: 0.426 (MFUE) -> 0.960 | Improved |
| use non-invasive fingerprint instead of watermark embedding | Adversarial Shifting vs SSL-WM · clean accuracy | CDA: SSL-WM average loss 0.41%, ours none | Improved |
| use anchor-guided group fingerprints | Adversarial Shifting · CIFAR-100 | AUC: 0.84 (MEA) -> 0.96 | Improved |
| use surrogate-free generation | Adversarial Shifting vs SSL-WM · CIFAR-100 | time: 681m -> 17m | Improved |
| prune 60-70% of parameters | Adversarial Shifting · GTSRB/CIFAR-100 (FTLL) | Matching Rate: > 0.9 | No clear effect |
| prune above 0.8 ratio | Adversarial Shifting · GTSRB/CIFAR-100 | Matching Rate | Hurt |
| hard-label model extraction (DFMS-HL) | Adversarial Shifting · GTSRB/CIFAR-100 | Matching Rate: < 0.3 | Hurt |
| embedding noise epsilon=0.1 | MEA · CIFAR-100 | Matching Rate: 0.43 | Hurt |
| embedding noise | Adversarial Shifting · CIFAR-100 | Matching Rate: > 0.8 | No clear effect |
| StolenEncoder extraction | Adversarial Shifting · GTSRB | Matching Rate: 0.6 | Mixed |
Closest papers by meaning
- SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models (CVPR 2025 · similarity 0.86)
- IPRemover: A Generative Model Inversion Attack against Deep Neural Network Fingerprinting and Watermarking (AAAI 2024 · similarity 0.85)
- United We Stand, Divided We Fall: Fingerprinting Deep Neural Networks via Adversarial Trajectories (NeurIPS 2024 · similarity 0.85)
- OmniMark: Efficient and Scalable Latent Diffusion Model Fingerprinting (AAAI 2025 · similarity 0.84)
- Fingerprinting Denoising Diffusion Probabilistic Models (CVPR 2025 · similarity 0.84)
- Intersecting-Boundary-Sensitive Fingerprinting for Tampering Detection of DNN Models (ICML 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.
