Reproduce “Disrupting Model Merging: A Parameter-Level Defense Without Sacrificing Accuracy” 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
Wei Junhao, Yu Zhe, Sakuma Jun
Researchers from RIKEN AIP and the Institute of Science Tokyo introduce PaRaMS, a parameter-level defense that proactively disrupts unauthorized machine learning model merging. The method preserves the protected model's original performance while causing significant degradation, such as a 65.16% accuracy drop in image classification, if merged without permission.
Key insights
- Model merging methods exploit the tendency of finetuned models, originating from the same pretrained checkpoint, to reside in a 'shared basin' within the loss landscape.
- Functionally equivalent parameter transformations can sufficiently perturb a protected model's parameter space to move it out of this shared basin relative to other models, thus disrupting merging capability without altering standalone performance.
- Combining both MLP parameter rearrangement and attention module scaling is crucial for achieving broad applicability and strong defense across diverse architectures and merging techniques.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.