[Paper Review] Evaluating Machine Unlearning via Epistemic Uncertainty

Alexander Becker, Thomas Liebig|arXiv (Cornell University)|Aug 23, 2022

Adversarial Robustness in Machine Learning4 citations

TL;DR

This paper introduces a novel evaluation metric for machine unlearning based on epistemic uncertainty, which quantifies how well a model has forgotten sensitive data without requiring retraining. The metric is theoretically grounded, efficient to compute, and outperforms adversarial attacks and retraining comparisons by capturing uncertainty changes linked to forgetting, especially in models updated toward a retrained solution.

ABSTRACT

There has been a growing interest in Machine Unlearning recently, primarily due to legal requirements such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act. Thus, multiple approaches were presented to remove the influence of specific target data points from a trained model. However, when evaluating the success of unlearning, current approaches either use adversarial attacks or compare their results to the optimal solution, which usually incorporates retraining from scratch. We argue that both ways are insufficient in practice. In this work, we present an evaluation metric for Machine Unlearning algorithms based on epistemic uncertainty. This is the first definition of a general evaluation metric for Machine Unlearning to our best knowledge.

Motivation & Objective

• To address the lack of a general, practical, and theoretically sound evaluation metric for machine unlearning algorithms.
• To overcome the limitations of existing evaluation methods, such as reliance on retraining (infeasible in practice) or adversarial attacks (inconclusive and expensive).
• To develop a metric rooted in epistemic uncertainty that reflects cognitive and information-theoretic principles of forgetting.
• To provide a computationally efficient upper bound for the metric, enabling scalability to larger models and datasets.
• To empirically validate the metric across diverse unlearning approaches and assess its sensitivity to parameter update directions.

Proposed method

• Proposes an efficacy metric based on epistemic uncertainty, measuring the reduction in model confidence for forgotten data points.
• Derives the metric from information-theoretic principles, linking uncertainty reduction to effective forgetting.
• Introduces a theoretical upper bound for the metric that preserves distributional relationships across forgetting percentages, enabling faster computation.
• Applies the metric to three unlearning methods: retraining, Fisher Forgetting, and Amnesiac Unlearning, comparing their efficacy.
• Uses membership inference attacks (MIAs) as a benchmark to validate the metric’s alignment with empirical unlearning success.
• Employs log-log plots to analyze the relationship between efficacy and MIA success rates, revealing model behavior under different forgetting strategies.

Experimental results

Research questions

• RQ1 Can epistemic uncertainty serve as a reliable and efficient proxy for measuring the success of machine unlearning?
• RQ2 How does the proposed metric compare to adversarial evaluation and retraining-based benchmarks in capturing actual forgetting?
• RQ3 Does the metric’s theoretical upper bound preserve meaningful distributional properties across different forgetting levels?
• RQ4 How do different unlearning strategies (e.g., retraining vs. parameter convergence to initial model) affect the metric’s behavior?
• RQ5 Is there a consistent relationship between the efficacy metric and the success rate of membership inference attacks?

Key findings

• The epistemic uncertainty-based efficacy metric successfully captures forgetting behavior and correlates with the success of membership inference attacks, validating its empirical relevance.
• The metric’s theoretical upper bound preserves the distribution of efficacy values across forgetting percentages, enabling efficient computation without loss of fidelity.
• Retraining and Fisher Forgetting show decreasing efficacy with increasing forgetting, supporting the hypothesis that unlearning toward a retrained model improves performance.
• Amnesiac Unlearning violates both hypotheses, as efficacy increases with more data points forgotten, due to parameter updates converging toward the initial model rather than the retrained one.
• The results demonstrate that the direction of parameter updates fundamentally affects unlearning outcomes, implying that evaluation metrics must be tailored to unlearning strategy.
• The study reveals that accuracy drops and adversarial attacks failing are necessary but insufficient indicators of successful unlearning, underscoring the need for uncertainty-based metrics.