[1]刘佳美,孙 涵,林 磊.基于伪标签的可防御稳定网络[J].计算机技术与发展,2022,32(06):34-38.[doi:10. 3969 / j. issn. 1673-629X. 2022. 06. 006]
　LIU Jia-mei,SUN Han,LIN Lei.Pseudo-label Based Defensible Stable Network[J].,2022,32(06):34-38.[doi:10. 3969 / j. issn. 1673-629X. 2022. 06. 006]
点击复制

基于伪标签的可防御稳定网络()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年06期
页码:
34-38
栏目:
大数据分析与挖掘
出版日期:
2022-06-10

文章信息/Info

Title:
Pseudo-label Based Defensible Stable Network
文章编号:
1673-629X(2022)06-0034-05
作者:
刘佳美; 孙 涵; 林 磊
南京航空航天大学 计算机科学与技术学院 / 人工智能学院,江苏 南京 211106
Author(s):
LIU Jia-mei; SUN Han; LIN Lei
School of Computer Science and Technology / Artificial Intelligence,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China
关键词:
域自适应; 聚类算法; 伪标签; 平均教师模型; 主动防御
Keywords:
domain adaptation; clustering algorithm; pseudo label; mean teacher model; active defense
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 06. 006
摘要:
针对域自适应问题中无法较好地同时提升模型迁移能力和防御攻击能力导致其在目标域中不稳定且易受攻击的问题,提出了一种基于伪标签的可防御稳定网络。 在条件域对抗网络的框架下,首先通过高斯混合模型对经过预训练输出的源域特征和目标域特征进行共同聚类,得到基于类别概率的软伪标签来引入更为可靠的目标域信息,以拉近两域之间的距离;接着将源域和目标域数据输入学生网络和教师网络,教师网络参数根据历史上学生网络参数通过指数移动平均方法迭代更新,通过约束特征的类内一致性以减轻错误的伪标签带来的不利影响;与此同时,采用主动防御的思想,在训练中增加源域的对抗样本,使模型学习到更鲁棒的特征,提高其在目标域数据对对抗攻击的防御能力。 在 Office-31 数据集上的实验结果表明,所提出的基于伪标签的可防御稳定网络能够有效提高模型的迁移能力和防御能力,从两个不同的方面提高了网络的鲁棒性。
Abstract:
Aiming at the problem that the migration ability and defense ability of the model cannot be improved at the same time in thedomain adaptation,which makes it unstable and vulnerable to attack in the target domain,a defensible and stable network based on pseudolabel is proposed. Under the framework of conditional domain adaptation network,firstly,the pre - trained source domain features andtarget domain features are co-clustered by Gaussian mixture model,and soft pseudo-label based on category probability is obtained to introduce more reliable target domain information,so as to shorten the distance between the two domains. Then the data are input into thestudent network and teacher network. According to the student network parameters,the teacher network parameters are updated by the exponential moving average to reduce the adverse effects of false labels. At the same time,using the idea of active defense,the model canlearn more robust features and improve the defense ability of the data in the target domain against the attack by adding the source domainadversarial samples in the training. Experimental results on Office - 31 show that the proposed algorithm can effectively improve themigration ability and defense ability of the model,and improve the robustness of the network from two different aspects.

相似文献/References:

[1]方杰 张结魁 周军.基于有向带权图的页面聚类算法研究[J].计算机技术与发展,2009,(09):49.
　FANG Jie,ZHANG Jie-kui,ZHOU Jun.Study on Page Clustering Algorithms Based on Weighted Directed Graph[J].,2009,(06):49.
[2]李静 陈立潮 成洪静 聂跃光.基于Delaunay三角网的CBDT聚类算法研究[J].计算机技术与发展,2009,(01):21.
　LI Jing,CHEN Li-chao,CHENG Hong-jing,et al.Study of Spatial Clustering Algorithm Based on Delaunay Triangulation[J].,2009,(06):21.
[3]金萍 王浩 宗瑜 李明楚.基于调整学习的聚类算法[J].计算机技术与发展,2009,(02):58.
　JIN Ping,WANG Hao,ZONG Yu,et al.Clustering Algorithm Based on Fine- Tuned Learning[J].,2009,(06):58.
[4]吴庆棋 林江云.基于聚类优化GMM提高说话人识别性能的研究[J].计算机技术与发展,2009,(04):35.
　WU Qing-qi,LIN Jiang-yun.A Study on GMM Optimization with Clustering for Improving Speaker Recognition[J].,2009,(06):35.
[5]耿筱媛 张燕平 闫屹.改进的K—means算法在电信客户细分中的应用[J].计算机技术与发展,2008,(05):163.
　GENG Xiao-yuan,ZHANG Yan-ping,YAN Yi.Application of Improved K - means Algorithm Subdivision of Telecom Clients[J].,2008,(06):163.
[6]朱永红.覆盖聚类算法的应用研究[J].计算机技术与发展,2007,(01):123.
　ZHU Yong-hong.Application Study on Covering Clustering Algorithm[J].,2007,(06):123.
[7]王鑫 王洪国 王珺 王金枝[].数据挖掘中聚类方法比较研究[J].计算机技术与发展,2006,(10):20.
　WANG Xin,WANG Hong-guo,WANG Jun,et al.Comparison of Clustering Methods in Data Mining[J].,2006,(06):20.
[8]赵慧 刘希玉 崔海青.网格聚类算法[J].计算机技术与发展,2010,(09):83.
　ZHAO Hui,LIU Xi-yu,CUI Hai-qing.Grid-Based Clustering Algorithm[J].,2010,(06):83.
[9]黄韬 刘胜辉 谭艳娜.基于k-means聚类算法的研究[J].计算机技术与发展,2011,(07):54.
　HUANG Tao,LIU Sheng-hui,TAN Yan-na.Research of Clustering Algorithm Based on K-means[J].,2011,(06):54.
[10]周峰 李龙澍.结合蚁群聚类算法的模糊C均值聚类[J].计算机技术与发展,2012,(07):45.
　ZHOU Feng,LI Long-shu.Fuzzy C Mean Clustering Combined Ant Colony Clustering Algorithm[J].,2012,(06):45.

更新日期/Last Update: 2022-06-10