Cross-Lingual Vulnerabilities in Machine Unlearning Evaluations
October 16, 2026
Research shows that machine unlearning often fails to achieve true information inaccessibility, as data can be recovered through alternative access paths like cross-lingual querying. Testing only on designated benchmark distributions overlooks significant privacy gaps in multilingual models.
HOW THIS AFFECTS YOU
●
researcherYou must evaluate unlearning via access-path generalization rather than just in-distribution forgetting.
●
policyCurrent unlearning standards may be insufficient for guaranteeing data privacy in multilingual deployments.