跳转至

文章背景与核心概要

在大语言模型(LLM)的机器遗忘研究中,传统方法通常假设预定义的“遗忘集”(Forget Set)与模型实际记忆的内容完全吻合。然而,在无法获取原始训练数据的现实隐私场景下,这种假设往往会失效,作者将这种差距定义为“遗忘集错位”(Forget-Set Misalignment)。本文深入剖析了由此产生的两种主要失败模式:遗忘不足(Under Unlearning)(遗忘集漏掉了实际记忆的数据,导致隐私泄露)以及知识外遗忘(Out-of-Knowledge Unlearning)(强迫模型遗忘它从未学过的内容,导致参数扰动和模型效用退化)。

为了解决这一根本性矛盾,本文作者团队提出了一种名为 CONfession-to-Forget-Set (CONFS) 的数据盲(Data-blind)框架。该框架通过诱导并形式化模型实际记忆的知识,构建出与模型内部状态高度对齐的遗忘集。实验证明,在合成基准、多模态以及真实世界基准测试中,CONFS 在多项指标上接近了黄金标准(Gold-standard)的性能,实现了出色的遗忘与效用平衡,并且在保护模型原有性能方面明显优于其他同类的数据盲遗忘集构建方法。


坦白你所知:大语言模型遗忘中“遗忘集与模型知识错位”问题研究

Confess What You Know: Forget-Set Misalignment with Model Knowledge in LLM Unlearning

📋 摘要

📋 Summary

This paper, accepted at EMNLP 2026, investigates the challenge of forget-set misalignment in Large Language Model (LLM) unlearning—specifically the discrepancy between a pre-defined forget set and what the model has actually memorized when original training data is inaccessible. The authors identify two primary failure modes: Under Unlearning (forget set misses memorized data, causing leakage) and Out-of-Knowledge Unlearning (forcing the model to forget things it never learned, leading to parameter perturbation and utility degradation). To resolve this, they introduce CONfession-to-Forget-Set (CONFS), a data-blind framework that builds model-aligned forget sets by eliciting the model's actual memorized knowledge.


📌 论文元数据

📌 Paper Metadata

  • arXiv ID: 2609.00605 [cs.LG]
  • 作者 (Authors): Miso Kim, Georu Lee, Seungwon Jeong, Woojin Lee
  • 提交时间 (Submitted): 2026年9月1日
  • 会议 (Conference): 已被 EMNLP 2026(主会)接收
  • 学科分类 (Subjects): 机器学习 (cs.LG); 人工智能 (cs.AI); 计算与语言 (cs.CL)

📖 摘要

大语言模型(LLM)的机器遗忘通常假设预定义的遗忘集与模型实际记忆的内容相匹配,但在无法访问原始训练数据的现实隐私场景中,这种假设经常被打破。我们将这种差距称为遗忘集错位(forget-set misalignment),并识别出两种情况:

📖 Abstract

Machine unlearning for large language models (LLMs) often assumes that a pre-defined forget set matches what the model has memorized, but this frequently breaks in realistic privacy settings where the original training data is inaccessible. We term this gap forget-set misalignment and identify two cases:

  1. 遗忘不足(Under Unlearning): 遗忘集遗漏了已记忆的信息,导致信息持续泄露。
  2. 知识外遗忘(Out-of-Knowledge Unlearning): 算法被迫去“遗忘”模型从未学习过的知识,从而扰动参数并降低模型效用。
  1. Under Unlearning: The forget set omits memorized information and leakage persists.
  2. Out-of-Knowledge Unlearning: The algorithm is driven to "forget" knowledge the model never learned, perturbing parameters and degrading utility.

通过梯度级分析,我们发现这些行为源于错位的遗忘目标,而不是特定的优化选择。为此,我们提出了 CONfession-to-Forget-Set (CONFS),这是一个数据盲框架,通过诱导并形式化模型的记忆知识来构建与模型对齐的遗忘集。在合成、多模态和真实世界的基准测试中,CONFS 在多项指标上接近黄金标准(Gold-standard)性能,实现了具有竞争力的遗忘-效用平衡,同时比其他数据盲遗忘集构建方法更好地保留了模型效用。

Using gradient-level analysis, we show these behaviors arise from misaligned unlearning targets rather than specific optimization choices. We then propose CONfession-to-Forget-Set (CONFS), a data-blind framework that constructs model-aligned forget sets by eliciting and formalizing the model's memorized knowledge. Across synthetic, multimodal, and real-world benchmarks, CONFS approaches Gold-standard performance on several metrics and achieves a competitive forgetting-utility balance, while preserving utility better than other data-blind forget-set constructions.


🔗 链接与资源