EAGER: Trustworthy and Verifiable Machine Unlearning in an Adversarial Environment

AI models are increasingly being trained on a huge amount of data. These models continuously need to remove selected data, or correct them to support a better accuracy or utility, and/or to comply with regulatory requirements such as the right to be forgotten/erasure/rectification. Machine Unlearning (MU) has emerged as a critical approach to support efficient removal of data, including privacy-sensitive or intellectual property information. However, MU approaches themselves may introduce new vulnerabilities that could be exploited to alter the resulting model behavior or to extract privacy-sensitive information. This exploratory project addresses these critical issues and aims to develop a trustworthy and verifiable MU framework to ensure the security and privacy of evolving and emergent AI systems. The project aims to gain a foundational understanding and will develop defensive techniques against attacks on MU in an integrated and efficient way. Creating trustworthy MU frameworks is critical toward achieving various goals articulated in the “2025 America’s Action Plan.” This project will explore novel conceptual and theoretical foundations of trustworthy MU by exploring: (1) foundational understanding of the MU process that leads to security and privacy risks; (2) building resilient defense against such threats; and (3) developing a holistic understanding of trustworthy MU. Toward this, the project will focus on the following research thrusts: (1) threat modeling and formal bounds on attack surface and game theory-based composite attack scenarios; (2) resilient defense against composite attack strategies; (3) verifiable continuous unlearning; and (4) an integrative framework for trustworthy MU. The project will hence contribute to the broader issue of how AI governance and regulatory compliance can be effectively achieved. The project team will broadly disseminate research results through conference publications and presentations/talks, and integrate them as course and/or lab modules for undergraduate and graduate courses. 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.