Towards Certified Unlearning for Deep Neural Networks

arXiv stat.ML / 4/23/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper addresses the gap between “certified unlearning” techniques that work well for convex models and the harder nonconvex setting of deep neural networks (DNNs).
It proposes several simple methods to extend certified unlearning to nonconvex objectives in DNN training.
To improve efficiency, the authors introduce an efficient computation approach using inverse Hessian approximation while preserving the model’s certification guarantees.
The work further broadens certification considerations to cover nonconvergent training and sequential unlearning requests occurring at different times.
Experiments on three real-world datasets show the proposed approach is effective and that certified unlearning provides benefits for DNNs.

Abstract

In the field of machine unlearning, certified unlearning has been extensively studied in convex machine learning models due to its high efficiency and strong theoretical guarantees. However, its application to deep neural networks (DNNs), known for their highly nonconvex nature, still poses challenges. To bridge the gap between certified unlearning and DNNs, we propose several simple techniques to extend certified unlearning methods to nonconvex objectives. To reduce the time complexity, we develop an efficient computation method by inverse Hessian approximation without compromising certification guarantees. In addition, we extend our discussion of certification to nonconvergence training and sequential unlearning, considering that real-world users can send unlearning requests at different time points. Extensive experiments on three real-world datasets demonstrate the efficacy of our method and the advantages of certified unlearning in DNNs.