时间序列分类模型的集成对抗训练防御方法

王璐瑶; 曹渊; 刘博涵; 曾恩; 刘坤; 夏元清

doi:10.16383/j.aas.c240050

时间序列分类模型的集成对抗训练防御方法

doi: 10.16383/j.aas.c240050 cstr: 32138.14.j.aas.c240050

王璐瑶^1,,
曹渊^1,,
刘博涵^1,,
曾恩^1,,
刘坤^1,,
夏元清^1,

1.
北京理工大学自动化学院自主智能无人系统全国重点实验室北京 100081

详细信息

作者简介:
王璐瑶：北京理工大学自动化学院硕士研究生. 主要研究方向为对抗攻击与防御, 时间序列, 网络安全. E-mail: luyaowangbit@gmail.com

曹渊：北京理工大学自动化学院博士研究生. 主要研究方向为对抗学习, 强化学习, 网络安全. E-mail: yuancao@bit.edu.cn

刘博涵：北京理工大学自动化学院硕士研究生. 主要研究方向为强化学习, 对抗训练, 网络安全. E-mail: bohanliu@bit.edu.cn

曾恩：北京理工大学自动化学院硕士研究生. 主要研究方向为深度学习, 对抗攻击与防御. E-mail: enzeng@bit.edu.cn

刘坤：北京理工大学自动化学院研究员. 主要研究方向为网络化控制理论与应用, 复杂网络安全. 本文通信作者. E-mail: kunliubit@bit.edu.cn

夏元清：北京理工大学自动化学院教授. 主要研究方向为云控制, 云数据中心优化调度管理, 智能交通, 模型预测控制, 自抗扰控制, 鲁棒控制, 复杂网络控制与安全, 网络化控制理论与应用, 飞行器控制和空天地一体化网络协同控制. E-mail: xia_yuanqing@bit.edu.cn

计量
- 文章访问数: 158
- HTML全文浏览量: 49
- PDF下载量: 22
- 被引次数: 0
出版历程
- 收稿日期: 2024-01-19
- 录用日期: 2024-07-23
- 网络出版日期: 2024-09-29

Ensemble Adversarial Training Defense for Time Series Classification Models

WANG Lu-Yao^1
,,
CAO Yuan^1
,,
LIU Bo-Han^1
,,
ZENG En^1
,,
LIU Kun^1
,,
XIA Yuan-Qing^1
,

1.
National Key Laboratory of Autonomous Intelligent Unmanned Systems, School of Automation, Beijing Institute of Technology, Beijing 100081

More Information

Author Bio:
WANG Lu-Yao　Master student at the School of Automation, Beijing Institute of Technology. Her research interest covers adversarial attacks and defense, time series, and cyber security

CAO Yuan　Ph.D. candidate at the School of Automation, Beijing Institute of Technology. His research interest covers adversarial learning, reinforcement learning, and cyber security

LIU Bo-Han　Master student at the School of Automation, Beijing Institute of Technology. His research interest covers reinforcement learning, adversarial training, and cyber security

ZENG En　Master student at the School of Automation, Beijing Institute of Technology. His research interest covers deep learning and adversarial attacks and defense

LIU Kun　Professor at the School of Automation, Beijing Institute of Technology. His research interest covers theory and applications of networked control, and security of complex networked systems. Corresponding author of this paper

XIA Yuan-Qing　Professor at the School of Automation, Beijing Institute of Technology. His research interest covers cloud control, cloud data center optimization scheduling and management, intelligent transportation, model predictive control, active disturbance rejection control, robust control, control and security of complex networked systems, theory and applications of networked control, and flight control and networked cooperative control for integration of space, air and earth

摘要

摘要: 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.
- 时间序列 /
- 对抗样本 /
- 对抗训练 /
- 模型鲁棒性
Abstract: Deep learning is one of the primary approaches to solve the time series classification (TSC) problems. However, TSC models based on deep learning are susceptible to adversarial attacks, leading to a significant decrease in model classification accuracy. This paper investigates the issue of defense against adversarial attacks for TSC models and designs an ensemble adversarial training (AT) defense method. Firstly, this paper designs an ensemble adversarial training defense framework for TSC models. It generates adversarial examples using various TSC models and attack methods, which are then used to train the target model. Secondly, in the process of generating adversarial examples, a local perturbation algorithm based on Shapelets is designed, and combined with the fast gradient sign method (FGSM) based on momentum iteration to achieve effective white-box attacks. Simultaneously, the adversarial attacks against surrogate models are designed using knowledge distillation (KD) and Wasserstein generative adversarial network (WGAN), achieving effective attacks when the attacker has no knowledge of the target model. Then, Kullback-Leibler (KL) divergence is incorporated into the adversarial training loss function to further enhance model robustness. Finally, the effectiveness of the proposed approach is validated on the multivariate time series classification dataset from the UEA archive.
- Time series /
- adversarial example /
- adversarial training (AT) /
- model robustness

HTML全文

图 1 时间序列分类算法模型示意图

Fig. 1 The time series classification algorithm models

下载: 全尺寸图片幻灯片

图 2 集成对抗训练防御整体框架图

Fig. 2 The overall framework diagram of the ensemble AT defense

下载: 全尺寸图片幻灯片

图 3 WGAN黑盒攻击模型结构图

Fig. 3 The WGAN black-box attack model structure

下载: 全尺寸图片幻灯片

图 4 针对ResCNN模型的不同攻击

Fig. 4 The different attacks against the ResCNN model

下载: 全尺寸图片幻灯片

图 5 针对不同模型的MI-FGSM攻击

Fig. 5 The MI-FGSM attacks against the different models

下载: 全尺寸图片幻灯片

图 6 使用Shapelets前后效果对比

Fig. 6 The effect comparison before and after using Shapelets

下载: 全尺寸图片幻灯片

图 7 黑盒对抗攻击在ArticularyWordRecognition上的结果

Fig. 7 The results of black-box adversarial attacks on the ArticularyWordRecognition

下载: 全尺寸图片幻灯片

表 1 分类模型训练参数设置

Table 1 The classification of the models for training parameter settings

参数名称	MultiTST	ResCNN	MLSTM-FCN	OmniScaleCNN
优化器	SGD	SGD	SGD	SGD
学习率	0.01	0.01	0.001	0.001
学习率衰减系数	0.1/30轮	0.1/30轮	无衰减	无衰减
动量系数	0.9	0.9	0.9	0.9
权重衰减系数	$2 \times10^{-4}$	$2 \times10^{-4}$	$2 \times10^{-4}$	$2 \times10^{-4}$
批次大小	128	128	256	256
训练轮次	100	100	300	300

下载: 导出CSV

表 2 分类模型准确率

Table 2 The accuracy rate of classification models

子数据集名称	MultiTST	ResCNN	MLSTM-FCN	OmniScaleCNN
ArticularyWordRecognition	0.980	0.973	0.983	0.983
AtrialFibrillation	0.333	0.200	0.333	0.267
BasicMotions	0.700	1.000	1.000	1.000
CharacterTrajectories	0.897	0.760	0.596	0.811
Cricket	0.750	0.986	0.986	1.000
EigenWorms	0.573	0.847	0.687	–
Epilepsy	0.804	0.971	0.848	0.783
ERing	0.904	0.848	0.919	0.889
EthanolConcentration	0.251	0.304	0.255	–
FaceDetection	0.550	0.527	0.540	0.518
FingerMovements	0.510	0.520	0.490	0.500
HandMovementDirection	0.635	0.297	0.405	0.189
Handwriting	0.279	0.224	0.492	0.626
Heartbeat	0.293	0.756	0.707	0.712
InsectWingbeat	0.588	0.100	0.100	–
JapaneseVowels	0.138	0.084	0.084	0.084
Libras	0.189	0.867	0.128	0.094
LSST	0.218	0.680	0.219	0.475
MotorImagery	0.460	0.530	0.510	–
NATOPS	0.656	0.928	0.728	0.639
PEMS-SF	0.168	0.723	0.127	0.127
PenDigits	0.121	0.986	0.196	0.106
PhonemeSpectra	0.023	0.308	0.051	0.056
RacketSports	0.651	0.842	0.572	0.770
SelfRegulationSCP1	0.693	0.823	0.195	0.563
SelfRegulationSCP2	0.528	0.494	0.506	0.500
SpokenArabicDigits	0.343	0.100	0.100	0.100
StandWalkJump	0.467	0.400	0.333	0.333
UWaveGestureLibrary	0.803	0.794	0.887	0.912
DuckDuckGeese	0.320	0.620	–	–

下载: 导出CSV

表 3 MI-FGSM的对抗攻击成功率

Table 3 The success rate of adversarial attacks based on MI-FGSM

子数据集名称	FGSM	PGD	MI-FGSM
ArticularyWordRecognition	$0.910 \pm 0.011$	$0.973 \pm 0.015$	$0.983 \pm 0.008$
BasicMotions	$0.680 \pm 0.010$	$0.820 \pm 0.019$	$0.870 \pm 0.017$
CharacterTrajectories	$0.596 \pm 0.063$	$0.616 \pm 0.118$	$0.622 \pm 0.046$
Cricket	$0.810 \pm 0.073$	$0.910 \pm 0.023$	$0.950 \pm 0.024$
EigenWorms	$0.710 \pm 0.023$	$0.840 \pm 0.011$	$0.790 \pm 0.054$
Epilepsy	$0.783 \pm 0.065$	$0.804 \pm 0.083$	$0.848 \pm 0.054$
ERing	$0.660 \pm 0.037$	$0.670 \pm 0.101$	$0.679 \pm 0.006$
EthanolConcentration	$0.180 \pm 0.100$	$0.250 \pm 0.064$	$0.270 \pm 0.005$
FaceDetection	$0.500 \pm 0.093$	$0.540 \pm 0.106$	$0.540 \pm 0.025$
FingerMovements	$0.490 \pm 0.017$	$0.500 \pm 0.003$	$0.570 \pm 0.094$
HandMovementDirection	$0.297 \pm 0.099$	$0.405 \pm 0.082$	$0.589 \pm 0.112$
Handwriting	$0.179 \pm 0.107$	$0.224 \pm 0.073$	$0.312 \pm 0.088$
Heartbeat	$0.293 \pm 0.091$	$0.326 \pm 0.117$	$0.342 \pm 0.061$
Libras	$0.697 \pm 0.037$	$0.867 \pm 0.060$	$0.910 \pm 0.115$
LSST	$0.720 \pm 0.024$	$0.860 \pm 0.019$	$0.820 \pm 0.084$
MotorImagery	$0.110 \pm 0.048$	$0.140 \pm 0.016$	$0.140 \pm 0.030$
NATOPS	$0.630 \pm 0.118$	$0.910 \pm 0.023$	$0.952 \pm 0.073$
PEMS-SF	$0.668 \pm 0.089$	$0.723 \pm 0.128$	$0.757 \pm 0.117$
PenDigits	$0.686 \pm 0.012$	$0.696 \pm 0.059$	$0.776 \pm 0.020$
PhonemeSpectra	$0.123 \pm 0.046$	$0.208 \pm 0.001$	$0.256 \pm 0.039$
RacketSports	$0.551 \pm 0.070$	$0.642 \pm 0.021$	$0.672 \pm 0.051$
SelfRegulationSCP1	$0.790 \pm 0.063$	$0.832 \pm 0.008$	$0.890 \pm 0.033$
SelfRegulationSCP2	$0.814 \pm 0.001$	$0.861 \pm 0.071$	$0.840 \pm 0.067$
StandWalkJump	$0.330 \pm 0.111$	$0.500 \pm 0.063$	$0.500 \pm 0.096$
UWaveGestureLibrary	$0.862 \pm 0.086$	$0.871 \pm 0.082$	$0.918 \pm 0.010$
DuckDuckGeese	$0.420 \pm 0.062$	$0.490 \pm 0.002$	$0.631 \pm 0.043$

下载: 导出CSV

表 4 基于Shapelets的局部扰动攻击实验结果

Table 4 Experimental results of the local perturbation attacks based on Shapelets

子数据集名称	MI-FGSM攻击		局部扰动攻击
子数据集名称	攻击成功率	扰动大小(MAE)	攻击成功率	扰动大小(MAE)
ArticularyWordRecognition	$0.983 \pm 0.007$	$0.425 \pm 0.022$	$0.923 \pm 0.025$	$0.088 \pm 0.007$
BasicMotions	$0.870 \pm 0.052$	$2.638 \pm 0.151$	$0.790 \pm 0.024$	$0.447 \pm 0.019$
CharacterTrajectories	$0.622 \pm 0.025$	$0.040 \pm 0.003$	$0.572 \pm 0.015$	$0.011 \pm 0.001$
Cricket	$0.950 \pm 0.034$	$0.947 \pm 0.020$	$0.895 \pm 0.040$	$0.247 \pm 0.016$
EigenWorms	$0.790 \pm 0.053$	$18.292 \pm 1.090\,$	$0.720 \pm 0.056$	$1.941 \pm 0.101$
Epilepsy	$0.848 \pm 0.030$	$0.359 \pm 0.020$	$0.808 \pm 0.022$	$0.098 \pm 0.006$
ERing	$0.679 \pm 0.021$	$0.490 \pm 0.035$	$0.600 \pm 0.019$	$0.220 \pm 0.013$
EthanolConcentration	$0.270 \pm 0.012$	$0.530 \pm 0.038$	$0.210 \pm 0.011$	$0.073 \pm 0.003$
FaceDetection	$0.540 \pm 0.027$	$2.411 \pm 0.126$	$0.490 \pm 0.020$	$0.824 \pm 0.050$
FingerMovements	$0.570 \pm 0.043$	$15.281 \pm 1.136\,$	$0.520 \pm 0.027$	$4.372 \pm 0.283$
HandMovementDirection	$0.589 \pm 0.041$	$9.308 \pm 0.703$	$0.518 \pm 0.035$	$4.721 \pm 0.308$
Handwriting	$0.312 \pm 0.020$	$1.019 \pm 0.070$	$0.297 \pm 0.024$	$0.241 \pm 0.012$
Heartbeat	$0.342 \pm 0.027$	$1.695 \pm 0.097$	$0.314 \pm 0.025$	$0.755 \pm 0.050$
Libras	$0.910 \pm 0.029$	$0.050 \pm 0.004$	$0.780 \pm 0.014$	$0.021 \pm 0.002$
LSST	$0.820 \pm 0.044$	$6.831 \pm 0.547$	$0.650 \pm 0.032$	$1.928 \pm 0.154$
MotorImagery	$0.140 \pm 0.009$	$25.570 \pm 1.946\,$	$0.080 \pm 0.005$	$7.880 \pm 0.596$
NATOPS	$0.952 \pm 0.057$	$0.321 \pm 0.021$	$0.910 \pm 0.025$	$0.754 \pm 0.050$
PEMS-SF	$0.757 \pm 0.031$	$0.050 \pm 0.004$	$0.742 \pm 0.019$	$0.025 \pm 0.002$
PenDigits	$0.776 \pm 0.029$	$4.940 \pm 0.303$	$0.658 \pm 0.036$	$2.132 \pm 0.135$
PhonemeSpectra	$0.256 \pm 0.015$	$8.019 \pm 0.623$	$0.256 \pm 0.020$	$3.894 \pm 0.215$
RacketSports	$0.672 \pm 0.054$	$3.294 \pm 0.208$	$0.647 \pm 0.025$	$1.944 \pm 0.131$
SelfRegulationSCP1	$0.890 \pm 0.062$	$7.573 \pm 0.506$	$0.740 \pm 0.056$	$3.491 \pm 0.216$
SelfRegulationSCP2	$0.840 \pm 0.049$	$4.885 \pm 0.293$	$0.790 \pm 0.032$	$1.994 \pm 0.160$
StandWalkJump	$0.500 \pm 0.029$	$0.579 \pm 0.041$	$0.500 \pm 0.022$	$0.384 \pm 0.031$
UWaveGestureLibrary	$0.918 \pm 0.030$	$0.383 \pm 0.018$	$0.722 \pm 0.028$	$0.175 \pm 0.009$
DuckDuckGeese	$0.631 \pm 0.050$	$14.750 \pm 1.089\,$	$0.597 \pm 0.046$	$4.579 \pm 0.302$

下载: 导出CSV

表 5 针对替代模型的黑盒对抗攻击实验结果

Table 5 Experimental results of the black-box adversarial attacks for the surrogate model

子数据集名称	知识蒸馏		黑盒对抗攻击
子数据集名称	原始模型准确率	替代模型准确率	攻击成功率	扰动大小
ArticularyWordRecognition	$0.973 \pm 0.014$	$0.966 \pm 0.016$	$0.923 \pm 0.025$	$0.352 \pm 0.026$
BasicMotions	$1.000 \pm 0.000$	$1.000 \pm 0.000$	$0.740 \pm 0.059$	$1.722 \pm 0.114$
CharacterTrajectories	$0.760 \pm 0.036$	$0.720 \pm 0.027$	$0.544 \pm 0.029$	$0.075 \pm 0.003$
Cricket	$0.986 \pm 0.011$	$0.937 \pm 0.018$	$0.825 \pm 0.033$	$1.474 \pm 0.112$
EigenWorms	$0.847 \pm 0.038$	$0.810 \pm 0.042$	$0.690 \pm 0.048$	$20.120 \pm 1.201\,$
Epilepsy	$0.971 \pm 0.011$	$0.359 \pm 0.012$	$0.790 \pm 0.037$	$0.419 \pm 0.021$
ERing	$0.848 \pm 0.033$	$0.794 \pm 0.031$	$0.600 \pm 0.048$	$0.575 \pm 0.021$
FaceDetection	$0.527 \pm 0.028$	$0.518 \pm 0.020$	$0.500 \pm 0.038$	$2.411 \pm 0.171$
FingerMovements	$0.520 \pm 0.028$	$0.507 \pm 0.035$	$0.670 \pm 0.032$	$15.281 \pm 1.176\,$
Heartbeat	$0.756 \pm 0.033$	$0.749 \pm 0.049$	$0.411 \pm 0.025$	$3.117 \pm 0.147$
Libras	$0.867 \pm 0.051$	$0.822 \pm 0.058$	$0.740 \pm 0.042$	$0.054 \pm 0.003$
LSST	$0.680 \pm 0.041$	$0.660 \pm 0.031$	$0.630 \pm 0.046$	$7.447 \pm 0.437$
MotorImagery	$0.530 \pm 0.019$	$0.440 \pm 0.033$	$0.120 \pm 0.008$	$27.430 \pm 1.302\,$
NATOPS	$0.928 \pm 0.039$	$0.733 \pm 0.041$	$0.850 \pm 0.061$	$2.120 \pm 0.093$
PEMS-SF	$0.723 \pm 0.023$	$0.711 \pm 0.028$	$0.714 \pm 0.041$	$0.072 \pm 0.002$
PenDigits	$0.986 \pm 0.004$	$0.981 \pm 0.004$	$0.410 \pm 0.024$	$5.131 \pm 0.381$
PhonemeSpectra	$0.308 \pm 0.023$	$0.308 \pm 0.021$	$0.211 \pm 0.020$	$7.914 \pm 0.513$
RacketSports	$0.842 \pm 0.041$	$0.790 \pm 0.057$	$0.600 \pm 0.048$	$3.721 \pm 0.191$
SelfRegulationSCP1	$0.823 \pm 0.045$	$0.819 \pm 0.057$	$0.670 \pm 0.037$	$8.780 \pm 0.512$
SelfRegulationSCP2	$0.494 \pm 0.037$	$0.317 \pm 0.018$	$0.660 \pm 0.021$	$5.130 \pm 0.391$
StandWalkJump	$0.400 \pm 0.033$	$0.320 \pm 0.029$	$0.500 \pm 0.027$	$0.584 \pm 0.023$
UWaveGestureLibrary	$0.794 \pm 0.039$	$0.788 \pm 0.034$	$0.625 \pm 0.041$	$0.693 \pm 0.045$
DuckDuckGeese	$0.620 \pm 0.041$	$0.570 \pm 0.048$	$0.490 \pm 0.035$	$12.750 \pm 0.821\,$

下载: 导出CSV

表 6 C&W攻击下集成不同攻击的防御结果

Table 6 The defense results of ensemble different attacks under C&W attacks

子数据集名称	原始模型ResCNN	白盒对抗训练	黑盒对抗训练	白盒 + 黑盒对抗训练
ArticularyWordRecognition	0.427	0.426	0.517	0.683
AtrialFibrillation	0.000	0.200	0.200	0.222
BasicMotions	0.250	0.550	0.573	0.625
CharacterTrajectories	0.060	0.417	0.448	0.537
Cricket	0.264	0.528	0.598	0.778
DuckDuckGeese	0.210	0.320	0.352	0.375
EigenWorms	0.111	0.222	0.222	0.256
Epilepsy	0.029	0.406	0.500	0.565
ERing	0.159	0.559	0.600	0.674
EthanolConcentration	0.002	0.100	0.150	0.200
FaceDetection	0.500	0.523	0.523	0.523
FingerMovements	0.470	0.490	0.490	0.490
HandMovementDirection	0.203	0.216	0.216	0.216
Handwriting	0.034	0.095	0.095	0.149
Heartbeat	0.210	0.332	0.486	0.567
InsectWingbeat	0.100	0.100	0.100	0.100
JapaneseVowels	0.073	0.352	0.483	0.576
Libras	0.083	0.217	0.413	0.667
LSST	0.003	0.050	0.050	0.050
MotorImagery	0.111	0.244	0.244	0.256
NATOPS	0.050	0.083	0.100	0.150
PEMS-SF	0.050	0.100	0.100	0.100
PenDigits	0.100	0.300	0.300	0.400
PhonemeSpectra	0.003	0.020	0.020	0.050
RacketSports	0.083	0.200	0.200	0.300
SelfRegulationSCP1	0.111	0.200	0.200	0.222
SelfRegulationSCP2	0.050	0.100	0.100	0.100
SpokenArabicDigits	0.020	0.030	0.030	0.050
StandWalkJump	0.050	0.100	0.100	0.100
UWaveGestureLibrary	0.083	0.200	0.200	0.300

下载: 导出CSV

表 7 C&W攻击下集成不同数量分类模型的防御结果

Table 7 The defense results of ensemble different numbers of classification models under C&W attacks

子数据集名称	单一模型	集成两个模型	集成三个模型	集成防御	集成防御 + KL散度	LSTM-FWED
ArticularyWordRecognition	0.683	0.837	0.913	0.931	0.977	0.931
AtrialFibrillation	0.222	0.333	0.333	0.333	0.333	0.333
BasicMotions	0.625	0.715	0.753	0.796	0.815	0.875
CharacterTrajectories	0.537	0.567	0.747	0.751	0.751	0.543
Cricket	0.778	0.811	0.861	0.880	0.880	0.628
DuckDuckGeese	0.375	0.350	0.410	0.540	0.560	0.480
EigenWorms	0.256	0.256	0.433	0.472	0.498	0.378
Epilepsy	0.565	0.657	0.696	0.746	0.746	0.622
ERing	0.674	0.696	0.763	0.815	0.825	0.714
EthanolConcentration	0.200	0.180	0.200	0.300	0.300	0.200
FaceDetection	0.523	0.523	0.523	0.523	0.523	0.545
FingerMovements	0.490	0.490	0.490	0.510	0.510	0.520
HandMovementDirection	0.216	0.216	0.216	0.267	0.311	0.247
Handwriting	0.149	0.155	0.171	0.171	0.171	0.095
Heartbeat	0.567	0.558	0.722	0.730	0.730	0.756
InsectWingbeat	0.100	0.100	0.100	0.100	0.100	0.100
JapaneseVowels	0.576	0.759	0.844	0.900	0.900	0.844
Libras	0.667	0.598	0.733	0.746	0.746	0.568
LSST	0.050	0.050	0.050	0.169	0.169	0.350
MotorImagery	0.256	0.244	0.300	0.311	0.311	0.433
NATOPS	0.517	0.517	0.722	0.783	0.783	0.560
PEMS-SF	0.301	0.202	0.579	0.588	0.611	0.473
PenDigits	0.590	0.699	0.865	0.913	0.930	0.753
PhonemeSpectra	0.059	0.059	0.059	0.044	0.044	0.059
RacketSports	0.737	0.724	0.757	0.803	0.803	0.837
SelfRegulationSCP1	0.357	0.416	0.498	0.536	0.536	0.618
SelfRegulationSCP2	0.311	0.299	0.311	0.539	0.562	0.493
SpokenArabicDigits	0.100	0.100	0.100	0.100	0.100	0.100
StandWalkJump	0.400	0.375	0.533	0.53	0.533	0.533
UWaveGestureLibrary	0.612	0.712	0.806	0.838	0.855	0.622

下载: 导出CSV

A1 多变量时间序列分类数据集UEA

A1 Multivariate time series classification dataset UEA

子数据集名称	训练样本数	测试样本数	维度	序列长度	类别数
ArticularyWordRecognition	275	300	9	144	25
AtrialFibrillation	15	15	2	640	3
BasicMotions	40	40	6	100	4
CharacterTrajectories	1 422	1 436	3	182	20
Cricket	108	72	6	1 197	12
DuckDuckGeese	50	50	1 345	270	5
EigenWorms	128	131	6	17 984	5
Epilepsy	137	138	3	206	4
EthanolConcentration	261	263	3	1 751	4
ERing	30	270	4	65	6
FaceDetection	5 890	3 524	144	62	2
FingerMovements	316	100	28	50	2
HandMovementDirection	160	74	10	400	4
Handwriting	150	850	3	152	26
Heartbeat	204	205	61	405	2
InsectWingbeat	30 000	20 000	200	30	10
JapaneseVowels	270	370	12	29	9
Libras	180	180	2	45	15
LSST	2 459	2 466	6	36	14
MotorImagery	278	100	64	3 000	2
NATOPS	180	180	24	51	6
PenDigits	7 494	3 498	2	8	10
PEMS-SF	267	173	963	144	7
Phoneme	3 315	3 353	11	217	39
RacketSports	151	152	6	30	4
SelfRegulationSCP1	268	293	6	896	2
SelfRegulationSCP2	200	180	7	1 152	2
SpokenArabicDigits	6 599	2 199	13	93	10
StandWalkJump	12	15	4	2 500	3
UWaveGestureLibrary	120	320	3	315	8

下载: 导出CSV

参考文献(36)

[1]	夏元清, 闫策, 王笑京, 宋向辉. 智能交通信息物理融合云控制系统. 自动化学报, 2019, 45(1): 132−142 Xia Yuan-Qing, Yan Ce, Wang Xiao-Jing, Song Xiang-Hui. Intelligent transportation cyber-physical cloud control systems. Acta Automatica Sinica, 2019, 45(1): 132−142
[2]	张淇瑞, 孟思琪, 王兰豪, 刘坤, 代伟. 隐蔽攻击下信息物理系统的安全输出反馈控制. 自动化学报, 2024, 50(7): 1363−1372 Zhang Qi-Rui, Meng Si-Qi, Wang Lan-Hao, Liu Kun, Dai Wei. Secure output-feedback control for cyber-physical systems under stealthy attacks. Acta Automatica Sinica, 2024, 50(7): 1363−1372
[3]	Zhang Z, Li W, Bao R, Harimoto K, Wu Y, Sun X. ASAT: Adaptively scaled adversarial training in time series. Neurocomputing, 2023, 522: 11−23 doi: 10.1016/j.neucom.2022.12.013
[4]	Bai T, Luo J, Zhao J, Wen B, Wang Q. Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv: 2102.01356, 2021.
[5]	Goodfellow I. Defense against the dark arts: An overview of adversarial example security research and future research directions. arXiv preprint arXiv: 1806.04169, 2018.
[6]	Goodfellow I J, Shlens J, Szegedy C. Explaining and harnessing adversarial examples. arXiv preprint arXiv: 1412.6572, 2014.
[7]	张耀元, 原继东, 刘海洋, 王志海, 赵培翔. 基于局部扰动的时间序列预测对抗攻击. 软件学报, 2024, 35 (11): 5210−5227 doi: 10.13328/j.cnki.jos.007056 Zhang Yao-Yuan, Yuan Ji-Dong, Liu Hai-Yang, Wang Zhi-Hai, Zhao Pei-Xiang. Adversarial attack of time series forecasting based on local perturbations. Journal of Software, 2024, 35 (11): 5210−5227 doi: 10.13328/j.cnki.jos.007056
[8]	Kurakin A, Goodfellow I J, Bengio S. Adversarial examples in the physical world. In: Proceedings of the International Conference on Learning Representations. Toulon, France: ICLR, 2017.
[9]	Madry A, Makelov A, Schmidt L, Tsipras D, Vladu A. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv: 1706.06083, 2017.
[10]	Liu H Y, Ge Z J, Zhou Z Y, Shang F H, Liu Y Y, Jiao L C. Gradient correction for white-box adversarial attacks. IEEE Transactions on Neural Networks and Learning Systems, DOI: 10.1109/TNNLS.2023.3315414
[11]	Carlini N, Wagner D. Towards evaluating the robustness of neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy. California, USA: IEEE, 2017. 39−57
[12]	Zhang H, Yu Y, Jiao J, Xing E, El Ghaoui L, Jordan M. Theoretically principled trade-off between robustness and accuracy. In: Proceedings of the International Conference on Machine Learning. California, USA: ICML, 2019. 7472−7482
[13]	Tramèr F, Kurakin A, Papernot N, Goodfellow I, Boneh D, McDaniel P. Ensemble adversarial training: Attacks and defenses. arXiv preprint arXiv: 1705.07204, 2017.
[14]	Yang W, Yuan J, Wang X, Zhao P. TSadv: Black-box adversarial attack on time series with local perturbations. Engineering Applications of Artificial Intelligence, 2022, 114: Article No. 105218 doi: 10.1016/j.engappai.2022.105218
[15]	Harford S, Karim F, Darabi H. Generating adversarial samples on multivariate time series using variational autoencoders. IEEE/CAA Journal of Automatica Sinica, 2021, 8(9): 1523−1538 doi: 10.1109/JAS.2021.1004108
[16]	Qi S, Chen J, Chen P, Wen P, Shan W, Xiong L. An effective WGAN-based anomaly detection model for IoT multivariate time series. In: Proceedings of Pacific-Asia Conference on Knowledge Discovery and Data Mining. Osaka, Japan: Springer Nature Switzerland, 2023. 80−91
[17]	陈晋音, 沈诗婧, 苏蒙蒙, 郑海斌, 熊晖. 车牌识别系统的黑盒对抗攻击. 自动化学报, 2021, 47(1): 121−135 Chen Jin-Yin, Shen Shi-Jing, Su Meng-Meng, Zheng Hai-Bin, Xiong Hui. Black-box adversarial attack on license plate recognition system. Acta Automatica Sinica, 2021, 47(1): 121−135
[18]	Baluja S, Fischer I. Adversarial transformation networks: Learning to generate adversarial examples. arXiv preprint arXiv: 1703.09387, 2017.
[19]	Xiao C, Li B, Zhu J Y, He W, Liu M, Song D. Generating adversarial examples with adversarial networks. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence. Stockholm, Sweden: AAAI, 2018. 3905−3911
[20]	Li J, Yang Y, Sun J S, Tomsovic K, Qi H. ConAML: Constrained adversarial machine learning for cyber-physical systems. In: Proceedings of the ACM Asia Conference on Computer and Communications Security. New York, USA: ACM, 2021. 52−66
[21]	Karim F, Majumdar S, Darabi H. Adversarial attacks on time series. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 43(10): 3309−3320
[22]	Qian Z, Huang K, Wang Q F, Zhang X Y. A survey of robust adversarial training in pattern recognition: Fundamental, theory, and methodologies. Pattern Recognition, 2022, 131: Article No. 108889 doi: 10.1016/j.patcog.2022.108889
[23]	Zhang H, Chen H, Xiao C, Gowal S, Stanforth R, Li B, et al. Towards stable and efficient training of verifiably robust neural networks. In: Proceedings of the International Conference on Learning Representations. New Orleans, USA: ICLR, 2019.
[24]	Shafahi A, Najibi M, Ghiasi M A, Xu Z, Dickerson J, Studer C, et al. Adversarial training for free! In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. Vancouver BC, Canada: Curran Associates Inc, 2019. 3358−3369
[25]	Guo X, Zhang R, Zheng Y, Mao Y. Robust regularization with adversarial labelling of perturbed samples. arXiv preprint arXiv: 2105.13745, 2021.
[26]	Liu X, Hsieh C J. Rob-GAN: Generator, discriminator, and adversarial attacker. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. California, USA: IEEE, 2019. 11234−11243
[27]	陈晋音, 吴长安, 郑海斌, 王巍, 温浩. 基于通用逆扰动的对抗攻击防御方法. 自动化学报, 2023, 49(10): 2172−2187 Chen Jin-Yin, Wu Chang-An, Zheng Hai-Bin, Wang Wei, Wen Hao. Universal inverse perturbation defense against adversarial attacks. Acta Automatica Sinica, 2023, 49(10): 2172−2187
[28]	Liu L, Park Y, Hoang T N, Hasson H, Huan J. Robust multivariate time-series forecasting: Adversarial attacks and defense mechanisms. In: Proceedings of the International Conference on Learning Representations. Kigali, Rwanda: ICLR, 2022.
[29]	Liu Y, Xu L, Yang S, Zhao D, Li X. Adversarial sample attacks and defenses based on LSTM-ED in industrial control systems. Computers and Security, 2024, 140: Article No. 103750
[30]	Khan M, Wang H, Ngueilbaye A, Elfatyany A. End-to-end multivariate time series classification via hybrid deep learning architectures. Personal and Ubiquitous Computing, 2023, 27(2): 177−191 doi: 10.1007/s00779-020-01447-7
[31]	Zou X, Wang Z, Li Q, Sheng W. Integration of residual network and convolutional neural network along with various activation functions and global pooling for time series classification. Neurocomputing, 2019, 367: 39−45 doi: 10.1016/j.neucom.2019.08.023
[32]	He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Nevada, USA: IEEE, 2016. 770−778
[33]	Tang W, Long G, Liu L, Zhou T, Blumenstein M, Jiang J. Omni-Scale CNNs: A simple and effective kernel size configuration for time series classification. In: Proceedings of the International Conference on Learning Representations. Kigali, Rwanda: ICLR, 2022.
[34]	Zerveas G, Jayaraman S, Patel D, Eickhoff C. A transformer-based framework for multivariate time series representation learning. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York, USA: Association for Computing Machinery, 2021. 2114−2124
[35]	Bagnall A, Dau H A, Lines J, Flynn M, Large J, Bostrom A, et al. The UEA multivariate time series classification archive. arXiv preprint arXiv: 1811.00075, 2018.
[36]	夏元清. 云控制系统及其面临的挑战. 自动化学报, 2016, 42(1): 1−12 Xia Yuan-Qing. Cloud control systems and their challenges. Acta Automatica Sinica, 2016, 42(1): 1−12