arXiv cs.AI· Kun Fang, Qinghua Tao, Junxu Liu, Yaxin Xiao, Yulin Jin, Qingqing Ye, Jian Sun, Haibo Hu·· 13 小时前AI 评分31
LOFT：在低维特征子空间中实现机器遗忘
Machine Unlearning in Low-Dimensional Feature Subspace
AI 导读
研究者提出 LOFT，一种在预训练模型的低维特征子空间中执行机器遗忘（MU）的方法，通过优化小型投影矩阵分离剩余数据与遗忘数据。该方法只需从预训练模型一次性提取特征，无需反复访问原始数据或更新整个模型，实验显示其计算开销显著更低且遗忘性能更优。该工作已被 NeurIPS 2026 接收，代码已公开。
正文 · 原文
Abstract:Machine Unlearning (MU) aims at removing the influence of specific data from a pretrained model while preserving performance on the remaining data. In this work, a novel perspective for MU is presented upon low-dimensional feature subspaces, which gives rise to the potentials of separating the remaining and forgetting data herein. This separability motivates our LOFT, a method that proceeds unlearning in a LOw-dimensional FeaTure subspace from the pretrained model through principal projections, which are optimized to maximally capture the information of the remaining data and meanwhile diminish that of the forgetting data. During unlearning, LOFT simply optimizes a small-size projection matrix that can be flexibly inserted into the pretrained model, requiring just one-shot feature fetching from the pretrained model. LOFT demonstrates that effective and efficient MU can be achieved in low-dimensional feature subspaces even without repeated raw-data access or updates to the entire pretrained model. Extensive experiments validate the significantly lower computational overhead and superior unlearning performance of LOFT across diverse models, datasets, tasks, and applications. Code is publicly available.
| Comments: | This work has been accepted by NeurIPS 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2601.22456 [cs.LG] |
| (or arXiv:2601.22456v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2601.22456 arXiv-issued DOI via DataCite |
Submission history
From: Kun Fang Dr. [view email]
[v1]
Fri, 30 Jan 2026 01:58:38 UTC (526 KB)
[v2]
Sun, 18 Oct 2026 07:45:22 UTC (650 KB)
[v3]
Wed, 21 Oct 2026 07:55:34 UTC (651 KB)
来源：arXiv cs.AI · arxiv.org