Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols

Yongwoo Kim, Sungmin Cha, Donghyun Kim · 2025

Read Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols by Yongwoo Kim, Sungmin Cha, Donghyun Kim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-06991.txt.

Description

Yongwoo Kim, Sungmin Cha, and Donghyun Kim question how machine unlearning is graded, in this 2025 paper. Unlearning should erase chosen data from a trained model while keeping the rest intact, for privacy or legal reasons. Most checks read only the model's output scores on small test sets. The authors warn that such grades can give a false sense of security.

Their wider audit inspects the model's inner representations at large scale. Leading methods either wrecked the representation quality wholesale or merely relabeled outputs while keeping the erased data's imprint inside. They add a tougher test where forgotten classes resemble later task classes, forcing real representational change. The authors offer their protocol as a stricter standard for the field.

Catalog record

CREATORYongwoo Kim, Sungmin Cha, Donghyun Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-06991

Browse the full library · About the Tower of Babel