Erased, but Not Gone: Output Forgetting Is Not True Forgetting

The Evaluation Gap in Machine Unlearning

Machine Unlearning (MU) aims to remove the influence of specific data points from a trained model. The field currently relies heavily on output-level metrics—such as low accuracy on the forget set and reduced logit-level membership inference—to certify success. This paper challenges the sufficiency of these metrics, arguing that they provide a misleadingly optimistic view of unlearning. By comparing unlearned models against a model retrained from scratch (the 'retraining-consistent' reference), the authors demonstrate that models can appear to have 'forgotten' at the output layer while retaining significant, structured information in their internal representation space.

A Stronger Evaluative Lens

The authors introduce a diagnostic framework based on retraining-consistent representation forgetting. While standard metrics only check if the final predictions are correct, this approach examines whether the internal representations (the penultimate layer) align with those of a model that never saw the forget data. The study reveals that current unlearning methods often fail this more rigorous test. Even when a model produces the 'correct' output, its internal geometry remains distinct from the retrained reference, indicating that the influence of the forget set has not been truly erased.

Structured Failure Modes

The research identifies that the discrepancy between current unlearning methods and the ideal retrained model is not random noise. Instead, it follows a predictable structure:

• Forget/Retain Asymmetry: Unlearning methods often show partial alignment with retraining on the forget set but exhibit significant, often negative, alignment on the retain set.
• Directional Mismatch: The shifts in representation space induced by unlearning do not match the shifts induced by retraining, particularly regarding how retain-set performance is preserved.
• Concentrated Residuals: Residual leakage is not diffuse; it is concentrated along specific directions related to the retraining process, suggesting that current methods leave behind structured traces of the forgotten data.