R.04 · 기억 거버넌스 · 2021
Machine Unlearning
Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, Nicolas Papernot · IEEE Symposium on Security and Privacy 2021
Summary
Once a model has been trained on someone’s data it may have memorized it, and removing that influence is notoriously hard. The authors introduce SISA training, which limits the influence of each data point during training so that unlearning a request becomes faster.
CommentaryRemoving what a model has absorbed is costly and partial; removing a record kept outside the model can be complete — and checked.