Perturb to Forget: Zero-Shot Machine Unlearning

Hongjie Jia, Yu Shi, Di Wu, Jun Bai, Jielong Yang, Xiaohui Tao, Yan Li

Keywords: Machine Unlearning, Model Inversion, Neuron Perturbation

Abstract: Machine unlearning seeks to remove the influence of specific data from trained models, a requirement increasingly critical under modern privacy regulations. Yet most existing approaches either depend on costly retraining or require access to the original dataset, which may be unavailable or restricted. We propose Inversion-Guided Neuron Perturbation (IGNP), a zero-shot framework that performs unlearning entirely without the original data. IGNP begins by synthesizing class-representative samples through a model inversion-inspired process, enabling analysis of how different parameters encode forget and retain classes. By contrasting these sensitivities, IGNP identifies parameters that are especially critical for encoding the forget class, while being less influential for retain classes. This strategy erases targeted knowledge with precision while preserving model utility. Extensive experiments on multiple benchmarks demonstrate that IGNP achieves complete forgetting with minimal accuracy loss, outperforms state-of-the-art zero-shot and data-dependent baselines, and provides strong resistance to membership inference and inversion attacks. These results establish IGNP as a practical and efficient solution for data-free unlearning in compliance-driven machine learning.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 6888