![[2602.17088] MeGU: Machine-Guided Unlearning with Target Feature Disentanglement](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17088] MeGU: Machine-Guided Unlearning with Target Feature Disentanglement
Summary
The paper presents MeGU, a novel framework for machine unlearning that addresses the challenge of effectively erasing target data while preserving model utility. It utilizes Multi-modal Large Language Models for concept-aware re-alignment to enhance unlearning efficiency.
Why It Matters
As data privacy concerns grow, the ability to effectively 'forget' specific data points in machine learning models becomes crucial. MeGU offers a solution to the trade-off between unlearning and model performance, making it relevant for applications requiring compliance with privacy regulations.
Key Takeaways
- MeGU addresses the limitations of existing unlearning methods by leveraging concept-aware re-alignment.
- The framework uses Multi-modal Large Language Models to guide the unlearning process effectively.
- It introduces a positive-negative feature noise pair to disentangle target concept influence during model finetuning.
- MeGU enables controlled forgetting, mitigating risks of under-unlearning and over-unlearning.
- The approach is significant for enhancing data privacy compliance in machine learning applications.
Computer Science > Machine Learning arXiv:2602.17088 (cs) [Submitted on 19 Feb 2026] Title:MeGU: Machine-Guided Unlearning with Target Feature Disentanglement Authors:Haoyu Wang, Zhuo Huang, Xiaolong Wang, Bo Han, Zhiwei Lin, Tongliang Liu View a PDF of the paper titled MeGU: Machine-Guided Unlearning with Target Feature Disentanglement, by Haoyu Wang and 5 other authors View PDF HTML (experimental) Abstract:The growing concern over training data privacy has elevated the "Right to be Forgotten" into a critical requirement, thereby raising the demand for effective Machine Unlearning. However, existing unlearning approaches commonly suffer from a fundamental trade-off: aggressively erasing the influence of target data often degrades model utility on retained data, while conservative strategies leave residual target information intact. In this work, the intrinsic representation properties learned during model pretraining are analyzed. It is demonstrated that semantic class concepts are entangled at the feature-pattern level, sharing associated features while preserving concept-specific discriminative components. This entanglement fundamentally limits the effectiveness of existing unlearning paradigms. Motivated by this insight, we propose Machine-Guided Unlearning (MeGU), a novel framework that guides unlearning through concept-aware re-alignment. Specifically, Multi-modal Large Language Models (MLLMs) are leveraged to explicitly determine re-alignment directions for target s...
