AI Unlearning Methods May Not Truly Erase Data

Machine unlearning (MU) aims to remove specific data from a trained model, typically evaluated by whether the model's outputs no longer reflect the "forgotten" data. However, this research challenges the sufficiency of output-level evaluation, proposing that models might still retain traces of unlearned data in their internal representations, even if their external behavior suggests forgetting. The study introduces "retraining-consistent representation forgetting" as a more rigorous benchmark, comparing unlearned models against models trained from scratch without the forget data. Across various unlearning methods, datasets, and models, the findings indicate that standard output-level metrics often overstate the success of unlearning. The analysis reveals that unlearned models frequently exhibit a structured mismatch in their representation space compared to truly retrained models. This mismatch is characterized by an asymmetry between forgotten and retained samples, directional inconsistencies, and concentrated residual discrepancies along retraining-related directions. This suggests that current MU techniques achieve "apparent forgetting" rather than a complete, retraining-consistent erasure of information.