Shadow Unlearning: a neural semantic method for faithful anonymous unlearning in LLMs

The researchers propose Shadow Unlearning, an approximate unlearning paradigm that performs machine unlearning on anonymized forget data without exposing PII, and implement it with the NSPU framework, avoiding the exposure of existing methods to membership inference attacks due to accessing the data to be deleted. Experiments on multiple LLMs and the MuFU forget set covering five domains show that NSPU achieves better unlearning effectiveness, preserves model utility, and is at least 10 times more computationally efficient than standard unlearning methods.