Unlearnability and Unlearning for Model Dememorization

A recent study published on arXiv (2605.11592) explores sophisticated techniques for model dememorization, such as availability poisoning (unlearnability) and machine unlearning, which seek to mitigate data exploitation in machine learning. Unlearnability incorporates subtle alterations into data prior to its release to diminish learnability, whereas unlearning eliminates information after training. The study reveals common weaknesses: both approaches experience shallow dememorization, resulting in misleading assertions about reduced data learnability or forgetting due to weight changes. Additionally, input perturbations can influence subsequent unlearning, and unlearning might unintentionally retrieve domain knowledge obscured by unlearnability. This research underscores the necessity for stronger protective measures.