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

Yongwoo Kim, Sungmin Cha, Donghyun Kim

Abstract

Machine unlearning is a process to remove specific data points from a trained model while maintaining performance on retain data, addressing privacy or legal requirements. Despite its importance, existing unlearning evaluations tend to focus on logit-based metrics (i.e., accuracy) under small-scale scenarios. We observe that this could lead to a false sense of security in unlearning approaches under real-world scenarios. In this paper, we conduct a new comprehensive evaluation that employs representation-based evaluations of the unlearned model under large-scale scenarios to verify whether the unlearning approaches genuinely eliminate the targeted forget data from the model's representation perspective. Our analysis reveals that current state-of-the-art unlearning approaches either completely degrade the representational quality of the unlearned model or merely modify the classifier (i.e., the last layer), thereby achieving superior logit-based evaluation metrics while maintaining significant representational similarity to the original model. Furthermore, we introduce a rigorous unlearning evaluation setup, in which the forgetting classes exhibit semantic similarity to downstream task classes, necessitating that feature representations diverge significantly from those of the original model, thus enabling a more rigorous evaluation from a representation perspective. We hope our benchmark serves as a standardized protocol for evaluating unlearning algorithms under realistic conditions.

Introduction

Machine unlearning is the process of removing the influence of specific data points from a trained model—a critical capability for complying with "the right to be forgotten" (e.g., GDPR) or purging biased data. Ideally, an unlearned model should be indistinguishable from a model that was never exposed to the "forget" data (the "retrained" gold standard).

However, current evaluation protocols are fundamentally flawed. They rely almost exclusively on logit-based metrics, such as accuracy on the forget set. This narrow focus is misleading: a model can easily "game" these metrics by modifying its final classification head to hide certain classes while leaving the underlying feature representations (the encoder) intact. If the encoder still "recognizes" forbidden data, the unlearning is merely superficial and potentially reversible.

Methodology: Beyond Logits

To address these shortcomings, we propose a holistic evaluation framework that shifts focus from what a model predicts (logits) to what it encodes (representations).

Comparison between the traditional evaluation framework (left) and the proposed representation-based framework (right).

Our framework introduces several key components:

  1. Representation-Based Metrics: We utilize Centered Kernel Alignment (CKA) to measure similarity between internal feature maps and k-Nearest Neighbors (k-NN) Transferability to evaluate how well unlearned features generalize to downstream tasks.
  2. Top Class-wise Forgetting Scenario: Standard benchmarks often use random classes for forgetting, which is too simple due to feature overlap. We introduce a scenario where models must forget classes semantically similar to downstream tasks, forcing the algorithm to genuinely alter the feature space.
  3. Unified Metric (H-LR): We introduce the Harmonic mean of Logit and Representation gaps (H-LR) to penalize models that achieve high accuracy by merely tweaking the classifier while failing to transform internal representations.

Experimental Setup

We evaluate unlearning at scale, moving beyond small datasets like CIFAR-10 to high-dimensional distributions:

  • Primary Training: ImageNet-1K (approx. 1.28 million images).
  • Downstream Evaluation: Office-Home, CUB-200-2011 (Birds), and DomainNet-126.
  • Architectures: ResNet-50, Swin-Transformer, and ConvNeXt.

Key Results and Insights

Our analysis reveals several critical issues with current state-of-the-art (SOTA) unlearning methods.

The Shortcut Learning Trap

Many unlearning algorithms take the path of least resistance. They modify the final layer to satisfy accuracy requirements but leave the "knowledge" in the encoder untouched. As shown in the figure below, while some methods achieve a small logit-based evaluation gap (appearing close to the retrained model), their representation-based metrics (CKA) reveal they remain highly similar to the original model, failing to truly "forget" at a foundational level.

Discrepancy between logit-based and representation-based metrics. While some methods show a small logit-based gap (left), CKA analysis (right) reveals they remain representationaly similar to the original model rather than the retrained ideal.

Representational Collapse vs. Semantic Sensitivity

We observed that some methods, like Gradient Ascent (GA), cause "catastrophic collapse" of the entire feature space, rendering the model useless for downstream tasks. Furthermore, in our "Top Class-wise" scenario, the gap between unlearned and retrained models widened significantly, proving that semantic overlap is a major "crutch" for current methods. Interestingly, simple Pseudo-Labeling (PL)—mapping forget-class data to semantically similar "retain" classes—often outperformed complex SOTA methods in maintaining representational quality while achieving forgetting.

Conclusion

Existing unlearning evaluations often provide a false sense of security. By focusing only on output logits, they fail to detect when sensitive information remains embedded in a model's internal feature space. Our proposed benchmark, utilizing representation-based metrics and semantically challenging scenarios, provides a more rigorous and realistic protocol for evaluating unlearning algorithms. Future work should extend this framework to generative models and address the computational cost of producing "retrained" gold-standard models.

Summary updated

Abstract

Machine unlearning is a process to remove specific data points from a trained model while maintaining the performance on retain data, addressing privacy or legal requirements. Despite its importance, existing unlearning evaluations tend to focus on logit-based metrics (i.e., accuracy) under small-scale scenarios. We observe that this could lead to a false sense of security in unlearning approaches under real-world scenarios. In this paper, we conduct a new comprehensive evaluation that employs representation-based evaluations of the unlearned model under large-scale scenarios to verify whether the unlearning approaches genuinely eliminate the targeted forget data from the model's representation perspective. Our analysis reveals that current state-of-the-art unlearning approaches either completely degrade the representational quality of the unlearned model or merely modify the classifier (i.e., the last layer), thereby achieving superior logit-based evaluation metrics while maintaining significant representational similarity to the original model. Furthermore, we introduce a rigorous unlearning evaluation setup, in which the forgetting classes exhibit semantic similarity to downstream task classes, necessitating that feature representations diverge significantly from those of the original model, thus enabling a more rigorous evaluation from a representation perspective. We hope our benchmark serves as a standardized protocol for evaluating unlearning algorithms under realistic conditions.
BibTeX
@article{kim2026are,
  author = {Yongwoo Kim and Sungmin Cha and Donghyun Kim},
  title = {{Are We Truly Forgetting? A Critical Re-examination of Machine Unlearning Evaluation Protocols}},
  journal = {Engineering Applications of AI},
  year = {2026},
  doi = {10.1016/j.engappai.2026.113785},
}
Figure