OFMU: Optimization-Driven Framework for Machine Unlearning
Status: Accepted, ICLR, 2026
OFMU introduces an optimization-driven approach to machine unlearning that removes targeted knowledge while preserving downstream model utility. Instead of relying on a scalarized retention-forgetting objective, it is derived from bi-level optimality conditions and offers convergence-rate guarantees with improved forgetting–utility trade-offs compared with prior methods.
Recommended citation: Asif, S., & Mohammadi Amiri, M. (2026). "OFMU: Optimization-Driven Framework for Machine Unlearning." ICLR 2026.
Download Paper