<ul data-eligibleForWebStory="true"><li>Some machine unlearning methods might be susceptible to prompt attacks as per a recent study.</li><li>Eight unlearning techniques were evaluated across different model families revealing vulnerabilities in ELM.</li><li>Methods like RMU and TAR showed robust unlearning capabilities.</li><li>Specific prompt attacks, such as using Hindi filler text, could recover up to 57.3% accuracy in ELM models.</li><li>Logit analysis confirmed that unlearned models generally do not hide knowledge by altering output formatting.</li><li>The study challenges existing assumptions about the effectiveness of unlearning methods.</li><li>There is a need for better evaluation frameworks to distinguish between true knowledge removal and superficial output suppression.</li><li>The study authors have made their evaluation framework publicly available for assessing prompting techniques in retrieving unlearning knowledge.</li></ul>

Prompt Attacks Reveal Superficial Knowledge Removal in Unlearning Methods

Discover more