Reproduce “T2S: A Rehearsal-Based Approach for Extraction-Resistant Model Watermarking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jian-Ping Mei, Weibin Zhang, Ao Yao, Tiantian Zhu, Jie Xiao
T2S tackles watermark transfer by simulating extraction during fine-tuning and using the simulated model’s watermark loss to update the target. Under CIFAR-10 hard-label Knockoff extraction, its stolen model reaches 97.56 ± 2.97% watermark success rate, versus 72.40 ± 6.35% for SSW-S. This result is specific to a setup where the simulated model shares the target’s architecture and training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.