CRII: SaTC: The Right to be Forgotten in Follow-ups of Machine Learning: When Privacy Meets Explanation and Efficiency

点击翻译按钮获取中文摘要

The larger volume of user data collected from various sources has led to the advancement of machine learning and the adoption of these machine learning models in many real-world applications. However, user data is often highly sensitive, and unauthorized data releases have sparked increased concerns about privacy. In response, recent regulations compel organizations to allow users to proactively remove their data from a system for increased privacy protection. However, deleting data from machine-learning models and systems is non-trivial. Further, law scholars have criticized the continued use of machine learning models trained on deleted data instances as it violates privacy. This project significantly contributes to theoretically and empirically understanding the risk of privacy leaks in the context of machine learning models. The project's broader significance and importance are that the developed algorithms guarantee users the right to have their data and the influence of data completely deleted in systems such as social media, healthcare, finance, etc. This project addresses new research problems related to machine unlearning. It investigates privacy leakage risk and validity of (1) model explanation, (2) model pruning, and (3) transfer learning when user data has been deleted and machine unlearning happens. This project also tackles the subsequent tasks of designing unlearning algorithms to eliminate the influence of forgotten data from explanations, pruning models, and transferring knowledge. Frequent deletion requests often encounter expensive costs, and existing approximate unlearning algorithms cannot be applied to the formulation of explanation, pruning, and transfer learning. The investigator is developing efficient removal algorithms to perform an update step of forgetting data on the explanation, pruned structure, and transfer learning. This project also features a pipeline to handle frequent data deletion requests in high-stakes domains where user information is especially sensitive. With these contributions, the project will catalyze research progress on the topic of the right to be forgotten and ease users’ concerns about privacy in machine learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.