Level Unlearning

Level unlearning in machine learning focuses on selectively removing specific information—be it individual data points, user attributes, or entire entities—from a trained model without complete retraining. Current research explores various unlearning approaches at different levels (instance, entity, attribute, client), often employing techniques like contrastive learning, graph-based pruning, and modifications to federated learning algorithms to achieve efficient and accurate forgetting. This field is crucial for addressing privacy concerns mandated by data regulations and enhancing the security and fairness of machine learning models across diverse applications, including medical imaging, recommender systems, and large language models.