基于L_(1)正则化的近似机器遗忘算法研究  

A study on approximate machine forgetting algorithm based on L_(1) regularization

作　　者：吕思蓓 郭骁 LV Sibei;GUO Xiao(School of Mathematics,Northwest University,Xi′an 710127,China)

机构地区：[1]西北大学数学学院,陕西西安710127

出　　处：《纯粹数学与应用数学》2024年第4期595-605,共11页Pure and Applied Mathematics

基　　金：国家自然科学基金(12301384).

摘　　要：随着数据所有者对于数据隐私的重视程度不断增加,如何高效完成用户的数据删除请求成为相关机构关注的热点,并据此衍生出一系列关于机器遗忘算法的研究.本文提出了基于L1正则化的CR模型,利用ADMM算法得到了影响函数的显式解,并通过一次牛顿更新法来实现精准高效的机器遗忘.真实数据实验表明,所提方法在分类能力和运行时间上优于再训练模型和L2-CR模型.相对于再训练模型,时间缩短了超过100倍;相对于L2正则化模型,准确率提高了5-7个百分点,达到了遗忘效率和模型效用的权衡.With the increasing emphasis on data privacy by data owners,efficiently handling user data deletion requests has become a focal point for relevant organizations.This has led to a series of studies on machine forgetting algorithms.This paper proposes a Certified Removal(CR)model based on L1 regularization.The explicit solution of the influence function is obtained using the ADMM algorithm,and precise and efficient machine forgetting is achieved through a one-time Newton update.Real data experiments demonstrate that the proposed method outperforms both retraining models and L2-CR models in terms of classification accuracy and runtime.Compared to retraining models,the time is reduced by over 100 times.In comparison to L2 regularization models,the accuracy is improved by 5-7 percentage points,striking a balance between forgetting efficiency and model utility.

关 键 词：机器遗忘 L1正则化 一次牛顿更新 ADMM算法 

分 类 号：O212[理学—概率论与数理统计]