Unlearning similar samples: Can machine unlearning do better?

Researchers evaluated the effectiveness of existing machine unlearning methods when the training set contains samples similar to the target to be unlearned, and found that most methods struggle to thoroughly eliminate the influence of the target samples on image and language models, and even the retraining-from-scratch baseline is no exception. The paper conducts experiments on four carefully constructed datasets, points out that existing methods mostly focus on removing the samples themselves rather than the samples' influence on the model, which deviates from the basic definition of machine unlearning, and explores possible directions for improving existing unlearning schemes.