Reproduce “Embedding Watermarks into Deep Neural Networks” 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
Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, Shin'ichi Satoh
Researchers at KDDI Research and National Institute of Informatics developed a framework for embedding digital watermarks into deep neural networks during training. The method successfully embeds 256-bit watermarks with no bit error rate while preserving model performance and exhibiting robustness against fine-tuning and up to 80% parameter pruning.
Key insights
- Leveraging DNN over-parameterization, watermarks can be embedded during training through a regularizer, maintaining original task performance.
- The method's effectiveness hinges on the choice of the embedding matrix 'X', with random projections ('X_random') best preserving model utility and parameter distributions.
- Watermarks embedded via this training-time regularizer exhibit strong robustness against common model alterations such as fine-tuning and aggressive parameter pruning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.