Reproduce “Model Unmerging: Making Your Models Unmergeable for Secure Model Sharing” 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

Zihao Wang, Enneng Yang, Lu Yin, Shiwei Liu, Li Shen

MergeLock introduces a proactive defense mechanism to render fine-tuned AI models unmergeable, aiming to protect intellectual property and mitigate privacy and safety risks in shared models. It achieves this by applying specific invertible transformations to the self-attention layers of Transformer models, making unauthorized merging attempts drastically degrade performance (e.g., 72-84% accuracy drop).

Key insights

  • Self-attention layers in Transformers offer superior properties for continuous, invertible transformations due to the lack of element-wise non-linearities between linear projections, unlike Feedforward Networks.
  • Disrupting the linear mode connectivity by moving model parameters to a distant, yet functionally equivalent, loss basin is a highly effective strategy to prevent successful model merging.
  • A combination of random, permutation, and diagonal scaling matrices provides a robust and complex transformation that is difficult to reverse or align without the secret key.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.