Reproduce “Amnesiac Machine Learning” 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
Laura Graves, Vineel Nagisetty, Vijay Ganesh
Researchers at the University of Waterloo developed two methods, 'Unlearning' and 'Amnesiac Unlearning,' for efficiently removing the influence of sensitive data from deep neural networks. These techniques demonstrated effectiveness in reducing model accuracy on forgotten data and resisting model inversion and membership inference attacks, offering practical compliance with data privacy regulations.
Key insights
- Simply retraining a model after data deletion does not fully eliminate the model's knowledge of sensitive data or prevent privacy attacks.
- Different unlearning strategies can be tailored for specific needs: 'Unlearning' is effective for removing larger amounts of learned data, such as entire classes, while 'Amnesiac Unlearning' is ideal for precise removal of individual data points.
- Implementing 'Amnesiac Unlearning' requires maintaining records of parameter updates during training, introducing storage overhead but enabling highly precise and efficient data removal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.