基于通用逆扰动的对抗攻击防御方法

陈晋音; 吴长安; 郑海斌; 王巍; 温浩

doi:10.16383/j.aas.c201077

基于通用逆扰动的对抗攻击防御方法

doi: 10.16383/j.aas.c201077

陈晋音^{1, 2,},
吴长安^2,,
郑海斌^2,,
王巍^3,,
温浩^4,

1.
浙江工业大学网络空间安全研究院杭州 310023
2.
浙江工业大学信息工程学院杭州 310023
3.
中国电子科技集团公司第三十六研究所嘉兴 314001
4.
重庆中科云从科技有限公司重庆 401120

基金项目: 国家自然科学基金(62072406), 浙江省自然科学基金(LY19F020025), 教育部产学合作协同育人项目资助

详细信息

作者简介:
陈晋音：浙江工业大学网络空间安全研究院和信息工程学院教授. 2009年获得浙江工业大学博士学位. 主要研究方向为人工智能安全, 图数据挖掘和进化计算. 本文通信作者.E-mail: chenjinyin@zjut.edu.cn

吴长安：浙江工业大学硕士研究生. 主要研究方向为深度学习, 计算机视觉, 对抗攻击和防御. E-mail: wuchangan@zjut.edu.cn

郑海斌：浙江工业大学信息工程学院博士研究生. 主要研究方向为深度学习, 人工智能安全, 对抗攻击和防御, 图像识别. E-mail: haibinzheng320@gmail.com

王巍：中国电子科技集团公司第三十六研究所研究员. 主要研究方向为无线通信分析, 网络安全. E-mail: wwzwh@163.com

温浩：重庆中科云从科技有限公司高级工程师. 主要研究方向为量子通信, 计算机通信网络与大规模人工智能计算. E-mail: wenhao@cloudwalk.com

计量
- 文章访问数: 1437
- HTML全文浏览量: 1270
- PDF下载量: 183
- 被引次数: 0
出版历程
- 收稿日期: 2020-12-28
- 录用日期: 2021-04-16
- 网络出版日期: 2021-06-01
- 刊出日期: 2023-10-24

Universal Inverse Perturbation Defense Against Adversarial Attacks

CHEN Jin-Yin^{1, 2
,},
WU Chang-An^2
,,
ZHENG Hai-Bin^2
,,
WANG Wei^3
,,
WEN Hao^4
,

1.
Institute of Cyberspace Security, Zhejiang University of Technology, Hangzhou 310023
2.
College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023
3.
The 36th Research Institute of China Electronics Technology Group Corporation, Jiaxing 314001
4.
Chongqing Zhongke Yuncong Technology Co., Ltd., Chongqing 401120

Funds: Supported by National Natural Science Foundation of China (62072406), Natural Science Foundation of Zhejiang Province (LY19F020025), and Ministry of Education Industry-University Cooperation Collaborative Education Project

More Information

Author Bio:
CHEN Jin-Yin　Professor at the Institute of Cyberspace Security and the College of Information Engineering, Zhejiang University of Technology. She received her Ph.D. degree from Zhejiang University of Technology in 2009. Her research interest covers artificial intelligence security, graph data mining, and evolutionary computing. Corresponding author of this paper

WU Chang-An　Master student at the College of Information Engineering, Zhejiang University of Technology. His research interest covers deep learning, computer vision, adversarial attack and defense

ZHENG Hai-Bin　Ph.D. candidate at the College of Information Engineering, Zhejiang University of Technology. His research interest covers deep learning, artificial intelligence security, adversarial attack and defense, and image recognition

WANG Wei　Researcher at the 36th Research Institute of China Electronics Technology Group Corporation. His research interest covers wireless communication analysis and network security

WEN Hao　Senior engineer at Chongqing Zhongke Yuncong Technology Co., Ltd.. His research interest covers guantum communication, computer communication networks, and large-scale artificial intelligence computing

摘要

摘要: 现有研究表明深度学习模型容易受到精心设计的对抗样本攻击, 从而导致模型给出错误的推理结果, 引发潜在的安全威胁. 已有较多有效的防御方法, 其中大多数针对特定攻击方法具有较好防御效果, 但由于实际应用中无法预知攻击者可能采用的攻击策略, 因此提出不依赖攻击方法的通用防御方法是一个挑战. 为此, 提出一种基于通用逆扰动(Universal inverse perturbation, UIP)的对抗样本防御方法, 通过学习原始数据集中的类相关主要特征, 生成通用逆扰动, 且UIP对数据样本和攻击方法都具有通用性, 即一个UIP可以实现对不同攻击方法作用于整个数据集得到的所有对抗样本进行防御. 此外, UIP通过强化良性样本的类相关重要特征实现对良性样本精度的无影响, 且生成UIP无需对抗样本的先验知识. 通过大量实验验证, 表明UIP在不同数据集、不同模型中对各类攻击方法都具备显著的防御效果, 且提升了模型对正常样本的分类性能.
- 深度学习 /
- 通用逆扰动 /
- 对抗样本 /
- 通用防御
Abstract: Existing studies have shown that deep learning models are vulnerable to carefully crafted adversarial sample, leading to wrong decision by the model, which will cause potential security threats. Many effective defense methods have been proposed, most of which have good defense effects against specific attack methods. However, since the possible strategies of attackers cannot be predicted in practical applications, it is a challenge to propose a general defense that does not rely on attack methods. This paper proposes a defense method based on universal inverse perturbation (UIP), which generates universal inverse perturbation by learning important features of classes in the original data. UIP is universal to data and attack methods, that is, one UIP can realize defense against all samples obtained by different attack methods acting on the entire data set. In addition, UIP can guarantee the accuracy of benign samples by enhancing the important characteristics of the benign samples, and the generation of UIP does not require prior knowledge of adversarial samples. Extensive experiments are carried out to testify that UIP has a significant defense effect against various attack methods in different data sets and different models, and the model＇s classification performance for normal samples is improved as well.
- Deep learning /
- universal inverse perturbation (UIP) /
- adversarial example /
- general defense

HTML全文

图 1 通用逆扰动防御方法框图

Fig. 1 The framework of UIPD method

下载: 全尺寸图片幻灯片

图 2 基于特征分布和决策边界的UIPD分析示意图

Fig. 2 The UIPD analysis based on feature distribution and decision boundary

下载: 全尺寸图片幻灯片

图 3 基于鲁棒安全边界的UIPD分析示意图

Fig. 3 The UIPD analysis based on robust security boundaries

下载: 全尺寸图片幻灯片

图 4 MNIST数据集中不同模型的 UIP 可视化图

Fig. 4 The UIP visualization of MNIST dataset in different models

下载: 全尺寸图片幻灯片

图 5 参数敏感性实验结果图

Fig. 5 The results of Parameter sensitivity experiment

下载: 全尺寸图片幻灯片

图 6 不同防御方法实施1000次防御的时间消耗

Fig. 6 The time cost in 1000 defenses of different defense methods

下载: 全尺寸图片幻灯片

图 7 UIPD对AP攻击的防御实验结果

Fig. 7 The results of UIPD against AP attacks

下载: 全尺寸图片幻灯片

A1 不同数据集和模型的UIP可视化图

A1 The UIP visualization of different datasets and models

下载: 全尺寸图片幻灯片

表 1 自行搭建的网络模型结构

Table 1 The network structure built by ourselves

网络层	M_CNN/F_CNN
Conv + ReLU	5 × 5 × 5
Max pooling	2 × 2
Conv + ReLU	5 × 5 × 64
Max pooling	2 × 2
Dense (Fully connected)	1024
Dropout	0.5
Dense (Fully connected)	10
Softmax	10

下载: 导出CSV

表 2 UIPD针对不同攻击方法的防御成功率(%)

Table 2 The defense success rate of UIPD against different attack methods (%)

DSR	MNIST			FMNIST		CIFAR-10	ImageNet
DSR	AlexNet	LeNet	M_CNN	AlexNet	F_CNN	VGG19	VGG19
良性样本识别准确率	92.34	95.71	90.45	89.01	87.42	79.55	89.00
FGSM^[8]	73.31	85.21	77.35	79.15	80.05	78.13	43.61
BIM^[18]	99.30	93.73	99.11	95.28	97.61	85.32	72.90
MI-FGSM^[9]	69.65	90.32	98.99	88.35	85.75	56.93	44.76
PGD^[17]	99.31	95.93	99.19	97.80	95.83	81.05	73.13
C&W^[19]	99.34	96.04	92.10	96.44	94.44	80.67	46.67
L-BFGS^[6]	98.58	70.12	67.79	66.35	71.75	68.69	31.36
JSMA^[10]	64.33	55.59	76.61	72.31	69.51	60.04	37.54
DeepFool^[20]	98.98	97.98	94.52	93.54	91.63	83.13	62.54
UAP^[15]	97.46	97.09	99.39	97.85	96.55	83.07	72.66
Boundary^[12]	93.63	94.38	95.72	92.67	91.88	76.21	68.45
ZOO^[11]	77.38	75.43	76.39	68.36	65.42	61.58	54.18
AGNA^[21]	75.69	76.40	81.60	64.80	72.14	62.10	55.70
AUNA^[21]	74.20	73.65	78.53	65.75	62.20	62.70	52.40
SPNA^[21]	92.10	88.35	89.17	77.58	74.26	72.90	60.30

下载: 导出CSV

表 3 UIPD针对不同数据样本的通用性(MNIST, M_CNN)

Table 3 The universality of UIPD for different examples (MNIST, M_CNN)

第1组		第2组		第3组		第4组
良性样本类标	置信度	(良性样本 + UIP)类标	置信度	对抗样本类标	置信度	(对抗样本 + UIP) 类标	置信度
0	1.000	0	1.000	5	0.5390	0	0.9804
1	1.000	1	1.000	8	0.4906	1	0.9848
2	1.000	2	1.000	1	0.5015	2	0.9841
3	1.000	3	1.000	7	0.5029	3	0.9549
4	1.000	4	1.000	9	0.5146	4	0.9761
5	1.000	5	1.000	3	0.5020	5	0.9442
6	1.000	6	1.000	4	0.5212	6	0.9760
7	1.000	7	1.000	3	0.5225	7	0.8960
8	1.000	8	1.000	6	0.5228	8	0.9420
9	1.000	9	1.000	7	0.5076	9	0.9796

下载: 导出CSV

表 4 不同防御方法针对基于梯度的攻击的防御效果比较

Table 4 The performance comparison of different defense methods against gradient-based attacks

		MNIST			FMNIST		CIFAR-10	ImageNet
		AlexNet	LeNet	M_CNN	AlexNet	F_CNN	VGG19	VGG19
平均ASR (%)		95.46	99.69	97.88	98.77	97.59	87.63	81.79
DSR (%)	resize1	78.24	74.32	81.82	79.84	77.24	69.38	47.83
	resize2	78.54	64.94	78.64	79.34	69.65	64.26	43.26
	rotate	76.66	80.54	84.74	77.63	61.46	72.49	42.49
	Distil-D	83.51	82.08	80.49	85.24	82.55	75.17	57.13
	Ens-D	87.19	88.03	85.24	87.71	83.21	77.46	58.34
	D-GAN	72.40	68.26	70.31	79.54	75.04	73.05	51.04
	GN	22.60	30.26	27.56	27.96	22.60	23.35	13.85
	DAE	84.54	85.25	85.68	86.94	80.21	75.85	59.31
	APE-GAN	83.40	80.71	82.36	84.10	79.45	72.15	57.88
	UIPD	88.92	86.89	87.45	87.77	83.91	78.23	59.91
Rconf	resize1	0.9231	0.9631	0.9424	0.8933	0.9384	0.6742	0.4442
	resize2	0.8931	0.9184	0.9642	0.9731	0.9473	0.7371	0.4341
	rotate	0.9042	0.8914	0.9274	0.9535	0.8144	0.6814	0.4152
	Distil-D	0.9221	0.9053	0.9162	0.9340	0.9278	0.6741	0.4528
	Ens-D	0.9623	0.9173	0.9686	0.9210	0.9331	0.7994	0.5029
	D-GAN	0.8739	0.8419	0.8829	0.9012	0.8981	0.7839	0.4290
	GN	0.1445	0.1742	0.2452	0.1631	0.1835	0.1255	0.0759
	DAE	0.9470	0.9346	0.9633	0.9420	0.9324	0.7782	0.5090
	APE-GAN	0.8964	0.9270	0.9425	0.8897	0.9015	0.6301	0.4749
	UIPD	0.9788	0.9463	0.9842	0.9642	0.9531	0.8141	0.5141

下载: 导出CSV

表 5 不同防御方法针对基于优化的攻击的防御效果比较

Table 5 The performance comparison of different defense methods against optimization-based attacks

		MNIST			FMNIST		CIFAR-10	ImageNet
		AlexNet	LeNet	M_CNN	AlexNet	F_CNN	VGG19	VGG19
平均ASR (%)		93.28	96.32	94.65	95.20	93.58	88.10	83.39
DSR (%)	resize1	78.65	70.62	79.09	74.37	66.54	65.31	38.28
	resize2	63.14	67.94	77.14	66.98	63.09	62.63	41.60
	rotate	76.62	72.19	71.84	66.75	64.42	65.60	42.67
	Distil-D	82.37	82.22	80.49	82.47	83.28	71.14	45.39
	Ens-D	86.97	83.03	85.24	83.41	82.50	74.29	47.85
	D-GAN	82.43	80.34	86.13	79.35	80.47	70.08	43.10
	GN	20.16	21.80	25.30	19.67	18.63	21.40	13.56
	DAE	83.66	84.17	86.88	82.40	83.66	74.30	51.61
	APE-GAN	82.46	85.01	85.14	81.80	82.50	73.80	49.28
	UIPD	87.92	85.22	87.54	83.70	83.91	75.38	52.91
Rconf	resize1	0.8513	0.8614	0.8460	0.7963	0.8324	0.6010	0.3742
	resize2	0.7814	0.8810	0.8655	0.8290	0.8475	0.6320	0.3800
	rotate	0.8519	0.8374	0.8319	0.8100	0.8040	0.6462	0.4058
	Distil-D	0.9141	0.8913	0.9033	0.9135	0.9200	0.7821	0.4528
	Ens-D	0.9515	0.9280	0.8720	0.8940	0.9011	0.8155	0.4788
	D-GAN	0.8539	0.8789	0.8829	0.8733	0.8820	0.7450	0.4390
	GN	0.1630	0.1920	0.2152	0.1761	0.1971	0.1450	0.0619
	DAE	0.9120	0.9290	0.9510	0.9420	0.9324	0.7782	0.5090
	APE-GAN	0.8964	0.9270	0.9425	0.8897	0.9015	0.6301	0.4749
	UIPD	0.9210	0.9340	0.9520	0.9512	0.9781	0.8051	0.5290

下载: 导出CSV

表 6 不同防御方法处理后良性样本的识别准确率 (%)

Table 6 The accuracy of benign examples after processing by different defense methods (%)

	MNIST			FMNIST		CIFAR-10	ImageNet
	AlexNet	LeNet	M_CNN	AlexNet	F_CNN	VGG19	VGG19
良性样本	92.34	95.71	90.45	89.01	87.42	79.55	89.00
resize1	92.27 (−0.07)	95.66 (−0.05)	90.47 (+0.02)	88.97 (−0.04)	87.38 (−0.04)	79.49 (−0.06)	88.98 (−0.02)
resize2	92.26 (−0.80)	95.68 (−0.30)	90.29 (−0.16)	88.71 (−0.30)	87.38 (−0.04)	79.48 (−0.07)	87.61 (−1.39)
rotate	92.31 (−0.03)	95.68 (−0.03)	90.39 (−0.06)	88.95 (−0.06)	87.40 (0.02)	79.53 (−0.02)	88.82 (−0.18)
Distil-D	90.00 (−2.34)	95.70 (−0.01)	90.02 (−0.43)	88.89 (−0.12)	86.72 (−0.70)	76.97 (−2.58)	87.85 (−1.15)
Ens-D	94.35 (+2.01)	96.15 (+0.44)	92.38 (+1.93)	89.13 (+0.12)	87.45 (+0.03)	80.13 (+0.58)	89.05 (+0.05)
D-GAN	92.08 (−0.26)	95.18 (−0.53)	90.04 (−0.41)	88.60 (−0.41)	87.13 (−0.29)	78.80 (−0.75)	87.83 (−1.17)
GN	22.54 (−69.80)	25.31 (−70.40)	33.58 (−56.87)	35.71 (−53.30)	28.92 (−58.59)	23.65 (−55.90)	17.13 (−71.87)
DAE	91.57 (−0.77)	95.28 (−0.43)	89.91 (−0.54)	88.13 (−0.88)	86.80 (−0.62)	79.46 (−0.09)	87.10 (−1.90)
APE-GAN	92.30 (−0.04)	95.68 (−0.03)	90.42 (−0.03)	89.00 (−0.01)	87.28 (−0.14)	79.49 (−0.06)	88.88 (−0.12)
UIPD	92.37 (+0.03)	95.96 (+0.25)	90.51 (+0.06)	89.15 (+0.14)	87.48 (+0.06)	79.61 (+0.06)	89.15 (+0.15)

下载: 导出CSV

A1 UIPD针对不同数据样本的通用性(FMNIST, F_CNN)

A1 The universality of UIPD for different examples (FMNIST, F_CNN)

第1组		第2组		第3组		第4组
良性样本类标	置信度	(良性样本 + UIP)类标	置信度	对抗样本类标	置信度	(对抗样本 + UIP) 类标	置信度
0	1.000	0	1.000	6	0.4531	0	0.9415
1	1.000	1	1.000	3	0.4714	1	0.8945
2	1.000	2	1.000	6	0.5641	2	0.9131
3	1.000	3	1.000	1	0.5103	3	0.9425
4	1.000	4	1.000	2	0.4831	4	0.8773
5	1.000	5	1.000	7	0.5422	5	0.9026
6	1.000	6	1.000	5	0.4864	6	0.8787
7	1.000	7	1.000	5	0.5144	7	0.8309
8	1.000	8	1.000	4	0.4781	8	8.9424
9	1.000	9	1.000	7	0.4961	9	0.8872

下载: 导出CSV

A2 UIPD针对不同数据样本的通用性(CIFAR-10, VGG19)

A2 The universality of UIPD for different examples (CIFAR-10, VGG19)

第1组		第2组		第3组		第4组
良性样本类标	置信度	(良性样本 + UIP)类标	置信度	对抗样本类标	置信度	(对抗样本 + UIP) 类标	置信度
飞机	1.000	飞机	1.000	船	0.4914	飞机	0.9331
汽车	1.000	汽车	1.000	卡车	0.5212	汽车	0.9131
鸟	1.000	鸟	1.000	马	0.5031	鸟	0.8913
猫	1.000	猫	1.000	狗	0.5041	猫	0.9043
鹿	1.000	鹿	1.000	鸟	0.5010	鹿	0.8831
狗	1.000	狗	1.000	马	0.5347	狗	0.9141
青蛙	1.000	青蛙	1.000	猫	0.5314	青蛙	0.8863
马	1.000	马	1.000	狗	0.4814	马	0.8947
船	1.000	船	1.000	飞机	0.5142	船	0.9251
卡车	1.000	卡车	1.000	飞机	0.4761	卡车	0.9529

下载: 导出CSV

A3 UIPD针对不同数据样本的通用性(ImageNet, VGG19)

A3 The universality of UIPD for different examples (ImageNet, VGG19)

第1组		第2组		第3组		第4组
良性样本类标	置信度	(良性样本 + UIP)类标	置信度	对抗样本类标	置信度	(对抗样本 + UIP) 类标	置信度
导弹	0.9425	导弹	0.9445	军装	0.5134	导弹	0.8942
步枪	0.9475	步枪	0.9525	航空母舰	0.4981	步枪	0.7342
军装	0.9825	军装	0.9925	防弹背心	0.5014	军装	0.8245
皮套	0.9652	皮套	0.9692	军装	0.4831	皮套	0.8074
航空母舰	0.9926	航空母舰	0.9926	灯塔	0.4788	航空母舰	0.8142
航天飞机	0.9652	航天飞机	0.9652	导弹	0.5101	航天飞机	0.7912
防弹背心	0.9256	防弹背心	0.9159	步枪	0.4698	防弹背心	0.8141
灯塔	0.9413	灯塔	0.9782	客机	0.5194	灯塔	0.7861
客机	0.9515	客机	0.9634	坦克	0.4983	客机	0.7134
坦克	0.9823	坦克	0.9782	灯塔	0.5310	坦克	0.7613

下载: 导出CSV

参考文献(55)

[1]	Goodfellow I, Bengio Y, Courville A. Deep Learning. Cambridge: The MIT Press, 2016. 24−45
[2]	Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe, Nevada, USA: ACM, 2012. 1097−1105
[3]	Sutskever I, Vinyals O, Le Q V. Sequence to sequence learning with neural networks. In: Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal, Canada: ACM, 2014. 3104−3112
[4]	袁文浩, 孙文珠, 夏斌, 欧世峰. 利用深度卷积神经网络提高未知噪声下的语音增强性能. 自动化学报, 2018, 44(4): 751-759 doi: 10.16383/j.aas.2018.c170001 Yuan Wen-Hao, Sun Wen-Zhu, Xia Bin, Ou Shi-Feng. Improving speech enhancement in unseen noise using deep convolutional neural network. Acta Automatica Sinica, 2018, 44(4): 751-759 doi: 10.16383/j.aas.2018.c170001
[5]	代伟, 柴天佑. 数据驱动的复杂磨矿过程运行优化控制方法. 自动化学报, 2014, 40(9): 2005-2014 Dai Wei, Chai Tian-You. Data-driven optimal operational control of complex grinding processes. Acta Automatica Sinica, 2014, 40(9): 2005-2014
[6]	Szegedy C, Zaremba W, Sutskever I, Bruna J, Erhan D, Goodfellow I J, et al. Intriguing properties of neural networks. In: Proceedings of the 2nd International Conference on Learning Representations. Banff, Canada: ICLR, 2014.
[7]	Akhtar N, Mian A. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 2018, 6: 14410-14430 doi: 10.1109/ACCESS.2018.2807385
[8]	Goodfellow I J, Shlens J, Szegedy C. Explaining and harnessing adversarial examples. In: Proceedings of the 3rd International Conference on Learning Representations. San Diego, USA: ICLR, 2015.
[9]	Dong Y P, Liao F Z, Pang T Y, Su H, Zhu J, Hu X L, et al. Boosting adversarial attacks with momentum. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 9185−9193
[10]	Papernot N, McDaniel P, Jha S, Fredrikson M, Celik Z B, Swami A. The limitations of deep learning in adversarial settings. In: Proceedings of the IEEE European Symposium on Security and Privacy (EuroS&P). Saarbruecken, Germany: IEEE, 2016. 372−387
[11]	Chen P Y, Zhang H, Sharma Y, Yi J F, Hsieh C J. ZOO: Zeroth order optimization based black-box attacks to deep neural networks without training substitute models. In: Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security. Dallas, USA: ACM, 2017. 15−26
[12]	Brendel W, Rauber J, Bethge M. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In: Proceedings of the 6th International Conference on Learning Representations. Vancouver, Canada: ICLR, 2018.
[13]	Xiao C W, Li B, Zhu J Y, He W, Liu M Y, Song D. Generating adversarial examples with adversarial networks. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence. Stockholm, Sweden: IJCAI, 2018. 3905−3911
[14]	Papernot N, McDaniel P, Goodfellow I. Transferability in machine learning: From phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv: 1605.07277, 2016.
[15]	Moosavi-Dezfooli S M, Fawzi A, Fawzi O, Frossard P. Universal adversarial perturbations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE, 2017. 86−94
[16]	Ilyas A, Santurkar S, Tsipras D, Engstrom L, Tran B, Mądry A. Adversarial examples are not bugs, they are features. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. Vancouver, Canada: ACM, 2019. Article No. 12
[17]	Madry A, Makelov A, Schmidt L, Tsipras D, Vladu A. Towards deep learning models resistant to adversarial attacks. In: Proceedings of the 6th International Conference on Learning Representations. Vancouver, Canada: ICLR, 2018.
[18]	Kurakin A, Goodfellow I, Bengio S. Adversarial examples in the physical world. In: Proceedings of the 5th International Conference on Learning Representations.Toulon, France: ICLR, 2017.
[19]	Carlini N, Wagner D. Towards evaluating the robustness of neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (SP). San Jose, USA: IEEE, 2017. 39−57
[20]	Moosavi-Dezfooli S M, Fawzi A, Frossard P. DeepFool: A simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 2574−2582
[21]	Rauber J, Brendel W, Bethge M. Foolbox: A python toolbox to benchmark the robustness of machine learning models. arXiv preprint arXiv: 1707.04131, 2017.
[22]	Su J W, Vargas D V, Sakurai K. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 2019, 23(5): 828-841 doi: 10.1109/TEVC.2019.2890858
[23]	Baluja S, Fischer I. Adversarial transformation networks: Learning to generate adversarial examples. arXiv preprint arXiv: 1703.09387, 2017.
[24]	Cisse M, Adi Y, Neverova N, Keshet J. Houdini: Fooling deep structured prediction models. arXiv preprint arXiv: 1707.05373, 2017.
[25]	Sarkar S, Bansal A, Mahbub U, Chellappa R. UPSET and ANGRI: Breaking high performance image classifiers. arXiv preprint arXiv: 1707.01159, 2017.
[26]	Brown T B, Mané D, Roy A, Abadi M, Gilmer J. Adversarial patch. arXiv preprint arXiv: 1712.09665, 2017.
[27]	Karmon D, Zoran D, Goldberg Y. LaVAN: Localized and visible adversarial noise. In: Proceedings of the 35th International Conference on Machine Learning. Stockholm, Sweden: ICML, 2018. 2512−2520
[28]	Liu A S, Liu X L, Fan J X, Ma Y Q, Zhang A L, Xie H Y, et al. Perceptual-sensitive GAN for generating adversarial patches. In: Proceedings of the AAAI Conference on Artificial Intelligence. Honolulu, USA: AAAI, 2019. 1028−1035
[29]	Liu A S, Wang J K, Liu X L, Cao B W, Zhang C Z, Yu H. Bias-based universal adversarial patch attack for automatic check-out. In: Proceedings of the 16th European Conference on Computer Vision. Glasgow, UK: Springer, 2020. 395−410
[30]	Xie C H, Wang J Y, Zhang Z S, Zhou Y Y, Xie L X, Yuille A. Adversarial examples for semantic segmentation and object detection. In: Proceedings of the International Conference on Computer Vision (ICCV). Venice, Italy: IEEE, 2017. 1378−1387
[31]	Song C B, He K, Lin J D, Wang L W, Hopcroft J E. Robust local features for improving the generalization of adversarial training. In: Proceedings of the 8th International Conference on Learning Representations. Addis Ababa, Ethiopia: ICLR, 2020.
[32]	Miyato T, Dai A M, Goodfellow I J. Adversarial training methods for semi-supervised text classification. In: Proceedings of the 5th International Conference on Learning Representations. Toulon, France: ICLR, 2017.
[33]	Zheng S, Song Y, Leung T, Goodfellow I. Improving the robustness of deep neural networks via stability training. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 4480−4488
[34]	Dziugaite G K, Ghahramani Z, Roy D M. A study of the effect of JPG compression on adversarial images. arXiv preprint arXiv: 1608.00853, 2016.
[35]	Das N, Shanbhogue M, Chen S T, Hohman F, Chen L, Kounavis M E, et al. Keeping the bad guys out: Protecting and vaccinating deep learning with JPEG compression. arXiv preprint arXiv: 1705.02900, 2017.
[36]	Luo Y, Boix X, Roig G, Poggio T, Zhao Q. Foveation-based mechanisms alleviate adversarial examples. arXiv preprint arXiv: 1511.06292, 2015.
[37]	Gu S X, Rigazio L. Towards deep neural network architectures robust to adversarial examples. In: Proceedings of the 3rd International Conference on Learning Representations. San Diego, USA: ICLR, 2015.
[38]	Rifai S, Vincent P, Muller X, Glorot X, Bengio Y. Contractive auto-encoders: Explicit invariance during feature extraction. In: Proceedings of the 28th International Conference on International Conference on Machine Learning. Bellevue, USA: ACM, 2011. 833−840
[39]	Ross A S, Doshi-Velez F. Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: Proceedings of the 32nd AAAI Conference on Artificial Intelligence. Menlo Park, CA, USA: AAAI, 2018. 1660−1669
[40]	Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. arXiv preprintarXiv: 1503.02531, 2015.
[41]	Papernot N, McDaniel P, Wu X, Jha S, Swami A. Distillation as a defense to adversarial perturbations against deep neural networks. In: Proceedings of the IEEE Symposium on Security and Privacy (SP). San Jose, USA: IEEE, 2016. 582−597
[42]	Nayebi A, Ganguli S. Biologically inspired protection of deep networks from adversarial attacks. arXiv preprint arXiv: 1703.09202, 2017.
[43]	Cisse M, Adi Y, Neverova N, Keshet J. Houdini: Fooling deep structured visual and speech recognition models with adversarial examples. In: Proceedings of Advances in Neural Information Processing Systems. 2017.
[44]	Gao J, Wang B L, Lin Z M, Xu W L, Qi T J. DeepCloak: Masking deep neural network models for robustness against adversarial samples. In: Proceedings of the 5th International Conference on Learning Representations. Toulon, France: ICLR, 2017.
[45]	Jin J, Dundar A, Culurciello E. Robust convolutional neural networks under adversarial noise. arXiv preprint arXiv: 1511.06306, 2015.
[46]	Sun Z, Ozay M, Okatani T. HyperNetworks with statistical filtering for defending adversarial examples. arXiv preprint arXiv: 1711.01791, 2017.
[47]	Akhtar N, Liu J, Mian A. Defense against universal adversarial perturbations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 3389−3398
[48]	Hlihor P, Volpi R, Malagò L. Evaluating the robustness of defense mechanisms based on autoencoder reconstructions against Carlini-Wagner adversarial attacks. In: Proceedings of the 3rd Northern Lights Deep Learning Workshop. Tromsø, Norway: NLDL, 2020. 1−6
[49]	孔锐, 蔡佳纯, 黄钢. 基于生成对抗网络的对抗攻击防御模型. 自动化学报, DOI: 10.16383/j.aas.c200033 Kong Rui, Cai Jia-Chun, Huang Gang. Defense to adversarial attack with generative adversarial network. Acta Automatica Sinica, DOI: 10.16383/j.aas.c200033
[50]	Samangouei P, Kabkab M, Chellappa R. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In: Proceedings of the 6th International Conference on Learning Representations. Vancouver, Canada: ICLR, 2018.
[51]	Lin W A, Balaji Y, Samangouei P, Chellappa R. Invert and defend: Model-based approximate inversion of generative adversarial networks for secure inference. arXiv preprintarXiv: 1911.10291, 2019.
[52]	Jin G Q, Shen S W, Zhang D M, Dai F, Zhang Y D. APE-GAN: Adversarial perturbation elimination with GAN. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Brighton, UK: IEEE, 2019. 3842−3846
[53]	Xu W L, Evans D, Qi Y J. Feature squeezing: Detecting adversarial examples in deep neural networks. In: Proceedings of the 25th Annual Network and Distributed System Security Symposium. San Diego, USA: NDSS, 2018.
[54]	Ju C, Bibaut A, Van Der Laan M. The relative performance of ensemble methods with deep convolutional neural networks for image classification. Journal of Applied Statistics, 2018, 45(15): 2800-2818 doi: 10.1080/02664763.2018.1441383
[55]	Kim B, Rudin C, Shah J. Latent Case Model: A Generative Approach for Case-Based Reasoning and Prototype Classification, MIT-CSAIL-TR-2014-011, MIT, Cambridge, USA, 2014.