TLDR: This paper introduces new criteria and an efficient tool (A-LiRA) to audit approximate machine unlearning methods, especially for differentially private AI models. It reveals that current unlearning techniques often fail to adequately protect the privacy of removed data and can inadvertently increase the privacy risks for the data that remains in the model, highlighting a critical gap in current privacy-preserving AI practices.
The “right to be forgotten” is a fundamental privacy principle gaining traction globally, empowering individuals to request the deletion of their personal data from organizations’ records. In the realm of artificial intelligence and machine learning, this right translates into “machine unlearning” – the process of removing the influence of specific data points from a trained model without having to retrain the entire model from scratch. This is particularly crucial for large, complex AI models where full retraining is prohibitively expensive and time-consuming.
While machine unlearning sounds like a straightforward solution to data deletion requests, a recent research paper titled “Auditing Approximate Machine Unlearning for Differentially Private Models” by Yuechun Gu, Jiajie He, and Keke Chen from the University of Maryland, Baltimore County, sheds light on a critical, often overlooked aspect of this process: the privacy of the data that remains in the model.
The Unseen Privacy Challenge
Existing machine unlearning methods primarily focus on ensuring that the removed data’s influence is erased. The common assumption has been that the privacy of the retained data remains unaffected. However, this paper challenges that assumption, especially when dealing with models designed to be “differentially private.” Differential privacy (DP) is a strong mathematical guarantee that ensures individual data points do not significantly alter the output of an algorithm, thereby protecting privacy. When a model is differentially private, there’s an agreed-upon privacy budget (epsilon, denoted as ϵ) that defines the maximum acceptable privacy risk for any data point.
The researchers highlight a phenomenon known as the “privacy onion effect,” where removing some data points can inadvertently increase the privacy risks of other, retained data points. This is akin to thinning a crowd – the fewer people there are, the easier it is to identify individuals. This paper is the first to explore how this effect impacts differentially private models under existing approximate machine unlearning methods.
New Criteria for Comprehensive Privacy Auditing
To address this gap, the authors propose a holistic approach to auditing machine unlearning algorithms. They introduce two new privacy criteria for differentially private models:
-
Criterion 1 (Unlearned Samples): The privacy risk of the unlearned samples should be significantly reduced to a safe, pre-defined level after the unlearning process. Ideally, their risk should approximate what it would be if the model had never seen them.
-
Criterion 2 (Retained Samples): The privacy risk of the retained samples must remain below the original differential privacy budget (ϵ) agreed upon for the model. Unlearning should not inadvertently expose these samples to higher privacy risks.
These criteria are measured using Membership Inference Attacks (MIAs), which determine whether a specific data point was part of the training dataset. A higher True Positive Rate (TPR) at a given False Positive Rate (FPR) indicates a higher privacy risk.
Introducing A-LiRA: An Efficient Auditing Tool
To make this auditing process practical and efficient, the researchers developed a novel Membership Inference Attack called A-LiRA (Augmentation-based Likelihood Ratio Attack). Traditional powerful MIAs, like Online-LiRA, are computationally very expensive, often requiring many GPU hours per sample. A-LiRA significantly reduces this cost by utilizing data augmentation and approximating probability distributions, achieving comparable accuracy to Online-LiRA with an 88.3% reduction in time. This efficiency makes it feasible to audit machine unlearning methods on a larger scale.
Concerning Findings: Current Methods Fall Short
The experimental findings, using A-LiRA, revealed some concerning truths about current approximate machine unlearning algorithms (such as SUNSHINE, SSD, and SalUn):
-
Unlearned Samples: While approximate unlearning generally reduced the privacy risk of unlearned samples in differentially private models, in non-DP models, a small fraction of unlearned samples paradoxically became more sensitive after removal. This suggests that unlearning isn’t always a perfect privacy shield even for the data it targets.
-
Retained Samples: More critically, the study found that existing approximate machine unlearning algorithms often inadvertently compromise the privacy of retained samples in differentially private models. This means that after unlearning, some retained data points exceeded the original privacy budget (ϵ), effectively breaking the differential privacy guarantee and user agreements. This “privacy onion effect” was consistently observed, with unlearning more sensitive samples leading to more retained samples having their privacy breached.
Also Read:
- Targeted Data Removal: A New Strategy for Unlearning in Large Language Models
- New Framework Addresses Data Forgetting in Federated Learning with Information Theory
The Path Forward
The research paper, available at arXiv:2508.18671, concludes that the current landscape of approximate machine unlearning methods is not sufficient for comprehensively protecting privacy, especially in the context of differentially private models. The findings underscore an urgent need for the development of new “differentially private unlearning algorithms” that can simultaneously reduce the privacy risk of unlearned samples and maintain the privacy guarantees for all retained data. The proposed auditing criteria and the efficient A-LiRA tool offer a robust framework for evaluating and guiding the development of such next-generation unlearning techniques.


