Sample Efficient Corrective Deep Unlearning

Aniket Sharma, Nidhi Hegde

Keywords: machine unlearning, corrective unlearning, poisoned data

Abstract: Machine unlearning enables machine learning models to selectively forget a subset of training data, ensuring compliance with privacy laws and allowing for the efficient removal of outdated or harmful data samples. Current machine unlearning algorithms are restricted to specific models or are applicable only to a subset of learning and unlearning settings, while requiring full knowledge of data points to unlearn. In this paper, we propose a sample efficient corrective deep unlearning algorithm that achieves competitive empirical performance across various unlearning settings without degrading model performance. Our experiments demonstrate that our algorithm achieves strong unlearning performance while requiring only a small computation budget and a small unlearning sample size, thus making it a viable solution for scalable and practical machine unlearning.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 22738