深度学习是解决时间序列分类(Time series classification,TSC)问题的主要途径之一.然而,基于深度学习的TSC模型易受到对抗样本攻击,从而导致模型分类准确率大幅度降低.为此,研究了TSC模型的对抗攻击防御问题,设计了集成对抗训练(Advers...深度学习是解决时间序列分类(Time series classification,TSC)问题的主要途径之一.然而,基于深度学习的TSC模型易受到对抗样本攻击,从而导致模型分类准确率大幅度降低.为此,研究了TSC模型的对抗攻击防御问题,设计了集成对抗训练(Adversarial training,AT)防御方法.首先,设计了一种针对TSC模型的集成对抗训练防御框架,通过多种TSC模型和攻击方式生成对抗样本,并用于训练目标模型.其次,在生成对抗样本的过程中,设计了基于Shapelets的局部扰动算法,并结合动量迭代的快速梯度符号法(Momentum iterative fast gradient sign method,MI-FGSM),实现了有效的白盒攻击.同时,使用知识蒸馏(Knowledge distillation,KD)和基于沃瑟斯坦距离的生成对抗网络(Wasserstein generative adversarial network,WGAN)设计了针对替代模型的黑盒对抗攻击方法,实现了攻击者对目标模型未知时的有效攻击.在此基础上,在对抗训练损失函数中添加Kullback-Leibler(KL)散度约束,进一步提升了模型鲁棒性.最后,在多变量时间序列分类数据集UEA上验证了所提方法的有效性.展开更多
As a category of recurrent neural networks,echo state networks(ESNs)have been the topic of in-depth investigations and extensive applications in a diverse array of fields,with spectacular triumphs achieved.Nevertheles...As a category of recurrent neural networks,echo state networks(ESNs)have been the topic of in-depth investigations and extensive applications in a diverse array of fields,with spectacular triumphs achieved.Nevertheless,the traditional ESN and the majority of its variants are devised in the light of the second-order statistical information of data(e.g.,variance and covariance),while more information is neglected.In the context of information theoretic learning,correntropy demonstrates the capacity to grab more information from data.Therefore,under the guidelines of the maximum correntropy criterion,this paper proposes a correntropy-based echo state network(CESN)in which the first-order and higher-order information of data is captured,promoting robustness to noise.Furthermore,an incremental learning algorithm for the CESN is presented,which has the expertise to update the CESN when new data arrives,eliminating the need to retrain the network from scratch.Finally,experiments on benchmark problems and comparisons with existing works are provided to verify the effectiveness and superiority of the proposed CESN.展开更多
基金supported in part by the National Natural Science Foundation of China(62176109,62476115)the Fundamental Research Funds for the Central Universities(lzujbky-2023-ey07,lzujbky-2023-it14)+1 种基金the Natural Science Foundation of Gansu Province(24JRRA488)the Supercomputing Center of Lanzhou University
文摘As a category of recurrent neural networks,echo state networks(ESNs)have been the topic of in-depth investigations and extensive applications in a diverse array of fields,with spectacular triumphs achieved.Nevertheless,the traditional ESN and the majority of its variants are devised in the light of the second-order statistical information of data(e.g.,variance and covariance),while more information is neglected.In the context of information theoretic learning,correntropy demonstrates the capacity to grab more information from data.Therefore,under the guidelines of the maximum correntropy criterion,this paper proposes a correntropy-based echo state network(CESN)in which the first-order and higher-order information of data is captured,promoting robustness to noise.Furthermore,an incremental learning algorithm for the CESN is presented,which has the expertise to update the CESN when new data arrives,eliminating the need to retrain the network from scratch.Finally,experiments on benchmark problems and comparisons with existing works are provided to verify the effectiveness and superiority of the proposed CESN.
