基于Lipschitz常数约束的胶囊网络鲁棒性增强  

Robustness Improvement of Capsule Networks Based on Lipschitz Constant Constraint

作　　者：魏鹏飞 杨旻[1] 张媛媛[1] WEI Peng-fei;YANG Min;ZHANG Yuan-yuan(School of Mathematics and Information Sciences,Yantai University,Yantai 264005,China)

机构地区：[1]烟台大学数学与信息科学学院,山东烟台264005

出　　处：《烟台大学学报（自然科学与工程版）》2023年第2期143-148,203,共7页Journal of Yantai University(Natural Science and Engineering Edition)

基　　金：国家自然科学基金资助项目(11771257,11901506);山东省自然科学基金资助项目(ZR2018MA008)。

摘　　要：为了提升胶囊神经网络的对抗鲁棒性,本文提出了基于Lipschitz常数约束的正则化方法,利用对抗训练方法来得到更加鲁棒的胶囊网络,将改进后的模型命名为Lips-CapsNet。相比较其它模型而言,Lips-CapsNet计算简便,不需要对胶囊网络的结构做出任何改变。实验结果显示,模型在MNIST、SVHN等数据集上实现了鲁棒性的提升,尤其是在Fashion-MNIST数据集上,在较强的攻击算法下,对抗样本的预测精度提升可达8%。To improve the adversarial robustness of capsule network,this paper proposes a regularization method using the Lipschitz constant constraint and utilizing adversarial training to get a more robustness model which named the model Lips-CapsNet.Compared with other robust models,Lips-CapsNet is computationally simple and does not need to change the structure of the original network.Experimental results show that the proposed model achieves improved robustness on MNIST and SVHN datasets,especially on the Fashion-MNIST dataset,where the prediction accuracy of the adversarial samples is improved up to 8%.

关 键 词：胶囊网络 对抗鲁棒性 LIPSCHITZ常数 正则化 

分 类 号：O244[理学—计算数学]