Reproduce “Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings” 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

Shahad Hardan, Darya Taratynova, Abdelmajid Essofi, Karthik Nandakumar, Mohammad Yaqub

Forget-MI introduces a machine unlearning method for multimodal medical data, effectively unlearning both unimodal and joint patient information while maintaining model utility. It achieves superior privacy against inference attacks, recording the lowest MIA scores among approximate methods, and reduces AUC and F1 scores on forgotten data by 0.221 and 0.305, respectively.

Key insights

  • The proposed combination of unimodal and multimodal unlearning losses with corresponding retention losses effectively balances the need to erase specific information while preserving overall model utility on retained data.
  • Strategic application of small noise perturbations to the forget set is crucial for comprehensive unlearning, as it helps create a 'neighborhood' around the data points to be forgotten, inducing broader shifts in model representations.
  • The optimal weighting of unlearning versus retention losses varies with the size of the forget set; for larger forgotten percentages, prioritizing retention becomes more critical to mitigate performance degradation on non-forgotten data.

Summary by alphaXiv.

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