TFR-Net: A New Approach to Continual Machine Unlearning

A recent study published on arXiv presents the Trajectory-guided Forget-Recover Network (TFR-Net), aimed at tackling issues in continual machine unlearning. The goal of machine unlearning is to remove the impact of sensitive data from a model. However, in practical applications, unlearning requests are frequent, leading to two primary challenges: first, an unlearning action may inadvertently cause previously forgotten knowledge to resurface by redistributing computations; second, repeated unlearning can diminish the model's capacity to retain useful information. TFR-Net monitors channel-level risks across requests, distinguishing between persistent and transient channels while suppressing only the persistent ones. It also enhances model capacity by reactivating dormant channels that significantly support retained utility and exhibit low forget risk. Recovery is permitted only if retained utility remains intact. The paper can be found on arXiv under the identifier 2608.03123.