基于并行ADMM的高维数据隐私保护  

High-dimensional Data Privacy Protection Based on Parallel ADMM

作　　者：王鹏飞 王逊 WANG Pengfei;WANG Xun(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212100)

机构地区：[1]江苏科技大学计算机学院,镇江212100

出　　处：《计算机与数字工程》2025年第2期517-522,共6页Computer & Digital Engineering

摘　　要：传统ADMM方法在面对组结构的高维数据时,存在收敛速度慢、识别能力低的问题。因此论文基于双层惩罚变量模型,使用跨特征的并行交替方向乘子法(PADMM)算法来求解,并使用优化算法Nadam来提高模型的收敛速度。考虑数据在进行迭代的时候会面临着隐私泄露的风险,因此在模型求解的过程中加入适量的噪声,通过对算法的输出结果进行扰动,从而达到对数据隐私保护的目的。实验表明,使用跨特征的PADMM求解模型时,在收敛速度和识别能力方面都要优于传统方式的ADMM。向模型加入适量噪声后,对比不同的程度的噪声,发现模型都能有较高的准确率。结论表明论文所提出的算法模型能够很好地识别具有组结构的高维数据,同时也能有效地对数据隐私进行保护。The traditional ADMM method suffers from slow convergence and low recognition capability when faced with high-dimensional data with group structure.This paper therefore uses the parallel alternating directional multiplier method(PAD⁃MM)algorithm based on a two-layer penalised variable model across features to solve it,and uses the optimisation algorithm Nad⁃am to improve the convergence speed of the model.Considering the risk of privacy leakage when the data is iterated,a moderate amount of noise is added to the model solving process to protect the privacy of the data by perturbing the output of the algorithm.Ex⁃periments have shown that the PADMM solving model using cross-features outperforms the traditional way of ADMM in terms of con⁃vergence speed and recognition capability.After adding a moderate amount of noise to the model,this paper compares different lev⁃els of noise and found that the model is able to achieve high accuracy.The conclusion shows that the algorithmic model proposed in this paper is able to identify high-dimensional data with group structure well,while also effectively protecting data privacy.

关 键 词：跨特征PADMM 隐私保护 高维数据 

分 类 号：TP309.2[自动化与计算机技术—计算机系统结构]