This paper introduces SAFER, a continual unlearning framework designed to address critical issues like knowledge erosion and forgetting reversal in repeated machine unlearning scenarios. It aims to maintain model utility and prevent the reactivation of forgotten information, making AI systems more reliable and privacy-compliant over their lifecycle.

This paper introduces UnSLU-BENCH, the first benchmark for machine unlearning in spoken language understanding (SLU), evaluating eight unlearning techniques across various datasets and models. It also proposes the Global Unlearning Metric (GUM) to comprehensively assess efficacy, utility, and efficiency in unlearning requests, particularly concerning speaker-specific data for privacy.