TrustErase: Auditable Instant Machine Unlearning with Passport-Embedded Representations

TrustErase is a novel framework for machine unlearning that employs passport-embedded representations, allowing for instant, modular, and auditable forgetting without the need for retraining or access to original data. By using cryptographic keys within parameter-efficient adaptation layers, it enables the removal of specific classes or datasets through simple deactivation. Evaluations on MNIST, CIFAR10, and CIFAR100 demonstrate that TrustErase achieves or surpasses benchmarks set by existing methods like DELETE, L2UL, and Boundary Shrink, thereby offering a new approach to privacy-compliant AI that enhances transparency and compliance.