Pretrained Language Models (LMs) memorize a vast amount of knowledge during initial pretraining, including information that may violate the privacy of personal lives and identities. Previous work addressing privacy issues for language models has mostly focused on data preprocessing and differential privacy methods, both requiring re-training the underlying LM. We propose knowledge unlearning as an alternative method to reduce privacy risks for LMs post hoc. We show that simply performing gradient ascent on target token sequences is effective at forgetting them with little to no degradation of general language modeling performances for larger LMs; it sometimes even substantially improves the underlying LM with just a few iterations. We also find that sequential unlearning is better than trying to unlearn all the data at once and that unlearning is highly dependent on which kind of data (domain) is forgotten. By showing comparisons with a previous data preprocessing method and a decoding method known to mitigate privacy risks for LMs, we show that unlearning can give a stronger empirical privacy guarantee in scenarios where the data vulnerable to extraction attacks are known a priori while being much more efficient and robust. We release the code and dataset needed to replicate our results at https://github.com/joeljang/knowledge-unlearning.
Dual-Space Smoothness for Robust and Balanced LLM Unlearning
School of Data Science, The Chinese University of Hong Kong, Shenzhen（香港中文大学（深圳）数据科学学院）；Department of Computer Science and Engineering, University of Notre Dame（圣母大学计算机科学与工程系）
CURE:Circuit-Aware Unlearning for LLM-based Recommendation
University of Massachusetts Lowell（马萨诸塞大学洛厄尔分校）；Zhejiang University（浙江大学）；The Chinese University of Hong Kong（香港中文大学）；National University of Singapore（新加坡国立大学）
Forgetting-MarI: LLM Unlearning via Marginal Information Regularization
Department of Mathematics, University of California Davis, USA（加州大学戴维斯分校数学系）；SLAC National Accelerator Laboratory, Stanford University, USA（斯坦福大学SLAC国家加速器实验室）；Department of Computer Science, University of California Davis, USA（加州大学戴维斯分校计算机科学系）
欧盟的《通用数据保护条例》（GDPR）或美国的《加州消费者隐私法》（CCPA）引入的法规包括关于\textit{right to be forgotten}的规定，要求行业应用程序从其系统中删除与个人相关的数据。在使用机器学习来基于用户数据构建模型的几个现实世界行业应用中，这样的授权要求在数据清理以及模型再训练方面都付出巨大的努力，同时确保模型不会由于数据的移除而在预测质量上恶化。因此，如果这些应用程序以非常高的频率接收此类请求，则连续删除数据和模型重新训练步骤不会扩展。最近，一些研究人员提出了\textit{Machine Unlearning}的想法来应对这一挑战。尽管这一任务非常重要，但在自然语言处理（NLP）任务中，机器遗忘领域还没有得到充分的探索。在本文中，我们探讨了各种GLUE任务\cite{Wang:18}，如QQP，SST和MNLI的Unlearning框架。我们提出了计算效率高的方法（SISA-FC和SISA-A）来执行\textit{guaranteed} Unlearning，与基线相比，在保持模型性能不变的同时，在内存（90-95\%），时间（100倍）和空间消耗（99\%）方面都有显着减少。
MUSE: Machine Unlearning Six-Way Evaluation for Language Models
University of Washington（华盛顿大学）；Princeton University（普林斯顿大学）；University of Southern California（南加州大学）；University of Chicago（芝加哥大学）；Google Research（谷歌研究院）
Textual Unlearning Gives a False Sense of Unlearning
The State Key Laboratory of Blockchain and Data Security, Zhejiang University, P. R. China（区块链与数据安全国家重点实验室，浙江大学，中华人民共和国）；School of Cyber Science and Technology, Zhejiang University, P. R. China（网络安全与技术学院，浙江大学，中华人民共和国）；ETH Zurich, Switzerland（苏黎世联邦理工学院，瑞士）
