Unlearning Tabular Data Without a "Forget Set''

Machine unlearning is the process of removing the influence of specific data from a trained model. This paper introduces Reload, an algorithm designed for unlearning in tabular data settings without requiring access to the "forget set'' -- the data to be removed. Reload begins with a trained model, then uses the remaining training data, and cached gradients from the last training iteration to enable effective unlearning without the need to retain sensitive data, offering stronger privacy guarantees than existing methods. Our experiments on tabular data using the TabNet architecture demonstrate both efficient item-level unlearning and an extension to feature unlearning, which removes sensitive attributes. These results highlight how Reload has the potential potential to improve privacy-preserving machine learning in the tabular setting.