In modern machine learning (ML), both public and proprietary data is used to train massive machine learning models, using large computational resources. However, this raises a natural privacy concern, as some users are not interested in participating in the training of these models. Accordingly, the right to be forgotten allows data producers (all of us), to request that the ML model will not be based on their data. A naive response of the learning agent is to delete the data of the user from the dataset and re-train the ML model from scratch. This is, of course, impractical, and so various methods for machine unlearning have been proposed, which attempt to be competitive with full re-train. However, these methods assume that the learning agent knows exactly the points in the dataset that were requested to be deleted by the user. In some settings, this is also impractical since the learning agent did not track (or didn’t even know in advance) what is the source of each data point. This project’s goal is to propose models for machine unlearning in the presence of unknown data origin.
