Machine Unlearning Under Retain-Forget Entanglement: When Forgetting Corrupts What You Want to Keep

arXiv·medium signal

Paper formalizes the 'retain-forget entanglement' problem in machine unlearning — closely related retained samples get corrupted when you try to forget nearby data. This isn't a theoretical edge case; it's the core failure mode when complying with data deletion requests (GDPR right-to-be-forgotten) on production models. Provides bounds on when unlearning is safe vs. when it silently degrades model quality.