TrustErase: Auditable Instant Machine Unlearning with Passport-Embedded Representations

A new machine unlearning framework called TrustErase enables artificial intelligence models to forget specific classes or datasets instantly without retraining or access to the original data, according to a paper posted to the arXiv preprint server on June 15, 2026 [1][2]. The framework, detailed in a submission to the Cryptography and Security section of arXiv, uses what its authors call passport-embedded representations — cryptographic keys concealed within parameter-efficient adaptation layers — to achieve what they describe as verifiable, data-free unlearning [2]. By deactivating these passports, a model can be made to forget targeted information immediately, without the computational cost of retraining or fine-tuning [2]. A singular value decomposition technique hides the passports inside model weights, making the unlearning actions transparent and provably compliant, the paper states [2]. Evaluations on the MNIST, CIFAR10, and CIFAR100 datasets showed that TrustErase matched or exceeded the performance of existing unlearning benchmarks, including DELETE, L2UL, and Boundary Shrink, while operating in a strictly data-free regime [2]. The authors argue the approach establishes a new paradigm for trustworthy and instantly forgettable AI systems [2]. The paper appeared on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month [6]. Founded in 1991, arXiv hosts papers across physics, mathematics, computer science, and related fields, and passed the two-million-article milestone by the end of 2021 [6]. Submissions are moderated but not peer-reviewed [6]. The TrustErase paper was surfaced through arXivLabs, a framework that allows community collaborators to develop and share experimental tools directly on the repository’s website [4]. Launched formally in 2020, arXivLabs provides a space for third-party projects — such as bibliographic explorers, code finders, and recommender systems — that add functionality for readers and authors [4][5]. Collaborators must adhere to arXiv’s values of openness, community, excellence, and user data privacy, and are granted only minimal, anonymized user data for the purpose of ensuring correct tool operation [4]. The arXivLabs program is currently on a temporary hiatus for new proposals while the development team focuses on modernizing arXiv’s infrastructure and moving systems to the cloud, though existing Labs and already-submitted proposals are unaffected [3].