Consent-Aware Machine Unlearning: A Verifiable Framework for Selective Influence Removal

Machine unlearning is the question of how a trained model can be made to act as if some of its training instances were never encountered without having to retrain the model from scratch. Interest in the topic has increased as privacy regulations such as GDPR have been introduced and people’s demands to have control over their data have grown. Removal methods, verification and proof-of-unlearning schemes and audit analyses exist in the literature. From the papers that we reviewed however, none combine the concepts of consent withdrawal, scoped removal, verification, and an auditable record in one workflow. We aim to propose a framework that we call Consent-Aware Machine Unlearning (CAMU) that initiates unlearning when a person withdraws consent, verifies that unlearning is actually taking place, and documents this process in an auditable record. We review recent proposals, compare them on the issues of consent handling and verifiability, and identify the gap created, before discussing three potential approaches to verifying removal: membership-inference auditing, influence-function auditing, and certified removal bounds. The design proposal is called CAMU. We do not report any experiments and implementation and evaluation is left for future research.