基于多维特征图知识蒸馏的对抗样本防御方法

doi:10.11959/j.issn.2096-109x.2022012

摘要/Abstract

摘要：

计算机视觉领域倾向使用深度神经网络完成识别任务，但对抗样本会导致网络决策异常。为了防御对抗样本，主流的方法是对模型进行对抗训练。对抗训练存在算力高、训练耗时长的缺点，其应用场景受限。提出一种基于知识蒸馏的对抗样本防御方法，将大型数据集学习到的防御经验复用到新的分类任务中。在蒸馏过程中，教师模型和学生模型结构一致，利用模型特征图向量作为媒介进行经验传递，并只使用干净样本训练。使用多维度特征图强化语义信息的表达，并且提出一种基于特征图的注意力机制，将特征依据重要程度赋予权重，增强蒸馏效果。所提算法在 Cifar100、Cifar10 等开源数据集上进行实验，使用FGSM（fast gradient sign method）、PGD（project gradient descent）、C＆amp;W（Carlini-Wagner attack）等算法进行白盒攻击，测试实验效果。所提方法在 Cifar10 干净样本的准确率超过对抗训练，接近模型在干净样本正常训练的准确率。在L2距离的PGD攻击下，所提方法效果接近对抗训练，显著高于正常训练。而且其学习成本小，即使添加注意力机制和多维度特征图等优化方案，算力需求依然远小于对抗训练，是一种轻量级的对抗防御方法。知识蒸馏作为一种神经网络学习方案，不仅能够学习正常样本的决策经验，而且可以提取鲁棒性特征。使用少量数据生成准确而鲁棒的模型，可以提高泛化能力，减少对抗训练带来的成本。

关键词: 深度学习, 对抗样本防御, 知识蒸馏, 多维度特征图

Abstract:

The neural network approach has been commonly used in computer vision tasks.However, adversarial examples are able to make a neural network generate a false prediction.Adversarial training has been shown to be an effective approach to defend against the impact of adversarial examples.Nevertheless, it requires high computing power and long training time thus limiting its application scenarios.An adversarial examples defense method based on knowledge distillation was proposed, reusing the defense experience from the large datasets to new classification tasks.During distillation, teacher model has the same structure as student model and the feature map vector was used to transfer experience, and clean samples were used for training.Multi-dimensional feature maps were utilized to enhance the semantic information.Furthermore, an attention mechanism based on feature map was proposed, which boosted the effect of distillation by assigning weights to features according to their importance.Experiments were conducted over cifar100 and cifar10 open-source dataset.And various white-box attack algorithms such as FGSM (fast gradient sign method), PGD (project gradient descent) and C＆amp;W (Carlini-Wagner attack) were applied to test the experimental results.The accuracy of the proposed method on Cifar10 clean samples exceeds that of adversarial training and is close to the accuracy of the model trained on clean samples.Under the PGD attack of L2 distance, the efficiency of the proposed method is close to that of adversarial training, which is significantly higher than that of normal training.Moreover, the proposed method is a light-weight adversarial defense method with low learning cost.The computing power requirement is far less than that of adversarial training even if optimization schemes such as attention mechanism and multi-dimensional feature map are added.Knowledge distillation can learn the decision-making experience of normal samples and extract robust features as a neural network learning scheme.It uses a small amount of data to generate accurate and robust models, improves generalization, and reduces the cost of adversarial training.

Key words: deep learning, adversarial examples defense, knowledge distillation, multi-dimensional feature maps

中图分类号:

TP309

邱宝琳, 易平. 基于多维特征图知识蒸馏的对抗样本防御方法[J]. 网络与信息安全学报, 2022, 8(2): 88-99.

Baolin QIU, Ping YI. Adversarial examples defense method based on multi-dimensional feature maps knowledge distillation[J]. Chinese Journal of Network and Information Security, 2022, 8(2): 88-99.

图/表 15

图1

图2

图3

图4

图5

图6

表1

干净样本和FGSM攻击在Cifar10上的准确率Table 1 Accuracy of the clean and FGSM attack on Cifar10"

训练方式	数据集	干净样本	FGSM
正常训练	$C i f a r 10$	$94 . 77 %$	$7 . 1 %$
对抗训练	$C i f a r 10$	$87 . 42 %$	$58 . 04 %$
Shafahi蒸馏	Cifar100→Cifar10	92.93%	46.13%
特征图蒸馏	Cifar100→Cifar10	85.67%	40.02%
频域蒸馏	Cifar100→Cifar10	87.81%	40.56%
Shafahi+mask	Cifar100→Cifar10	94.79%	61.30%
特征图+mask	$C i f a r 100 \to C i f a r 10$	$94 . 17 %$	$62 . 44 %$
频域+mask	$C i f a r 100 \to C i f a r 10$	$94 . 23 %$	$63 . 57 %$

表1

表2

PGD-L2攻击算法在Cifar10上的准确率Table 2 Accuracy of the PGD-L2 attack on Cifar10"

训练方式	数据集	PGD-5	PGD-10	PGD-20
正常训练	$C i f a r 10$	$0 . 01 %$	$0$	$0$
对抗训练	$C i f a r 10$	$45 . 57 %$	$31 . 06 %$	$24 . 95 %$
Shafahi蒸馏	Cifar100→Cifar10	13.15%	4.7%	2.35%
特征图蒸馏	Cifar100→Cifar10	33.27%	18.26%	12.91%
频域蒸馏	Cifar100→Cifar10	27.27%	13.07%	8.39%
Shafahi+mask	Cifar100→Cifar10	29.44%	17.71%	13.32%
特征图+mask	$C i f a r 100 \to C i f a r 10$	$37 . 41 %$	$28 . 63 %$	$23 . 49 %$
频域+mask	$C i f a r 100 \to C i f a r 10$	$39 . 94 %$	$30 . 44 %$	$25 . 14 %$

表2

表3

PGD-L∞攻击算法在Cifar10上的准确率Table 3 Accuracy of the PGD-L∞ attack on Cifar10"

训练方式	数据集	PGD-5	PGD-10	PGD-20
正常训练	$C i f a r 10$	$0 . 01 %$	$0$	$0$
对抗训练	$C i f a r 10$	$57 . 76 %$	$49 . 31 %$	$47 . 38 %$
Shafahi蒸馏	Cifar100→Cifar10	11.63%	4.78%	2.61%
特征图蒸馏	Cifar100→Cifar10	35.95%	26.15%	23.93%
频域蒸馏	Cifar100→Cifar10	30.49%	20.26%	17.46%
Shafahi+mask	Cifar100→Cifar10	25.95%	15.89%	12.25%
特征图+mask	$C i f a r 100 \to C i f a r 10$	$34 . 56 %$	$26 . 17 %$	$21 . 87 %$
频域+mask	$C i f a r 100 \to C i f a r 10$	$37 . 09 %$	$27 . 82 %$	$23 . 33 %$

表3

表4

C＆W攻击算法在Cifar10上的准确率Table 4 Accuracy of the C＆W attack on Cifar10"

训练方式	数据集	CW-5	CW-10	CW-20
正常训练	$C i f a r 10$	$0 . 015 %$	$0$	$0$
对抗训练	$C i f a r 10$	$53 . 84 %$	$46 . 54 %$	$45 . 27 %$
Shafahi蒸馏	Cifar100→Cifar10	8.71%	3.27%	1.77%
特征图蒸馏	Cifar100→Cifar10	32.85%	24%	22.46%
频域蒸馏	Cifar100→Cifar10	27.35%	17.98%	16.15%
Shafahi+mask	Cifar100→Cifar10	20.93%	13.19%	9.83%
特征图+mask	$C i f a r 100 \to C i f a r 10$	$28 . 44 %$	$20 . 39 %$	$16 . 64 %$
频域+mask	$C i f a r 100 \to C i f a r 10$	$32 . 17 %$	$23 . 44 %$	$19 . 66 %$

表4

图7

表5

不同算法在Cifar10的耗时情况Table 5 Time-consuming for Different Algorithms on Cifar10"

训练方式	数据集	耗时
正常训练	$C i f a r 10$	$19 m i n$
对抗训练	$C i f a r 10$	$5 . 5 h$
Shafahi蒸馏	Cifar100→Cifar10	24min
特征图蒸馏	Cifar100→Cifar10	29min
频域蒸馏	Cifar100→Cifar10	27min
Shafahi+mask	Cifar100→Cifar10	29min
特征图+mask	$C i f a r 100 \to C i f a r 10$	$35 m i n$
频域+mask	$C i f a r 100 \to C i f a r 10$	$33 m i n$

表5

表6

Shafahi+mask基于不同α在Cifar10上的分类准确率Table 6 Accuracy based on α with shafahi used mask on Cifar10"

权重α	Clean	FGSM	PGD-5	PGD-10	PGD-20
5×10^-1	94.65%	28.31%	0.55%	0.02%	0
5	94.80%	38.39%	2.15%	0.17%	0.05%
50×10¹	94.95%	52.78%	9.21%	2.61%	1.00%
5×10²	$94 . 79 %$	$61 . 3 %$	$25 . 95 %$	$15 . 89 %$	$12 . 25 %$
5×10³	93.99%	58.28%	27.41%	17.54%	13.22%

表6

表7

特征图+mask基于不同α在Cifar10上的分类准确率Table 7 Accuracy based on α with feature map used mask on Cifar10"

权重α	Clean	FGSM	PGD-5	PGD-10	PGD-20
5×10^-1	94.63%	33.29%	0.89%	0.05%	0.01%
5	95.12%	45.31%	3.38%	0.44%	0.09%
5×10¹	94.79%	58.92%	15.71%	6.21%	3.37%
$5 \times 1 0^{2}$	$94 . 17 %$	$62 . 44 %$	$34 . 56 %$	$26 . 17 %$	$21 . 87 %$
5×10³	91.67%	54.38%	29.78%	19.03%	13.32%

表7

表8

频域+mask基于不同α在Cifar10上的分类准确率Table 8 Accuracy based on α with frequency used mask on Cifar10"

权重α	Clean	FGSM	PGD-5	PGD-10	PGD-20
0.01	94.75%	43.68%	1.73%	0.12%	0.02%
0.1	95.03%	51.42%	7.87%	1.78%	0.43%
1	95.05%	62.78%	27.53%	17.40%	13.36%
$10$	$94 . 23 %$	$63 . 57 %$	$37 . 09 %$	$27 . 82 %$	$23 . 33 %$
20	93.20%	57.47%	33.34%	28.40%	25.92%
30	87.89%	42.49%	19.55%	15.11%	11.96%

表8

参考文献 29

[1]	SZEGEDY C , ZAREMBA W , SUTSKEVER I ,et al. Intriguing properties of neural networks[J]. arXiv preprint arXiv:1312.6199, 2013.
[2]	XU K D , ZHANG G Y , LIU S J ,et al. Adversarial T-shirt evading person detectors in a physical world[C]// Proceedings of Computer Vision - ECCV 2020. 2020: 665-681.
[3]	HE K M , ZHANG X Y , REN S Q ,et al. Deep residual learning for image recognition[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. 2016: 770-778.
[4]	SZEGEDY C , VANHOUCKE V , IOFFE S ,et al. Rethinking the inception architecture for computer vision[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2818-2826.
[5]	HINTON G E , VINYALS O , DEAN J . Distilling the knowledge in a neural network[J]. arXiv preprint arXiv:1503.02531, 2015.
[6]	ZHOU B L , KHOSLA A , LAPEDRIZA A ,et al. Learning deep features for discriminative localization[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2921-2929.
[7]	易平, 王科迪, 黄程 ,等. 人工智能对抗攻击研究综述[J]. 上海交通大学学报, 2018,52(10): 1298-1306.
	YI P , WANG K D , HUANG C ,et al. Adversarial attacks in artificial intelligence:a survey[J]. Journal of Shanghai Jiao Tong University, 2018,52(10): 1298-1306.
[8]	王科迪, 易平 . 人工智能对抗环境下的模型鲁棒性研究综述[J]. 信息安全学报, 2020,5(3): 13-22.
	WANG K D , YI P . A survey on model robustness under adversarial example[J]. Journal of Cyber Security, 2020,5(3): 13-22.
[9]	GOODFELLOW I J , SHLENS J , SZEGEDY C . Explaining and harnessing adversarial examples[J]. arXiv preprint arXiv:1412.6572, 2014.
[10]	KURAKIN A , GOODFELLOW I , BENGIO S . Adversarial machine learning at scale[J]. arXiv preprint arXiv:1611.01236, 2016.
[11]	DONG Y P , LIAO F Z , PANG T Y ,et al. Boosting adversarial attacks with momentum[C]// Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018: 9185-9193.
[12]	MADRY A , MAKELOV A , SCHMIDT L ,et al. Towards deep learning models resistant to adversarial attacks[J]. arXiv preprint arXiv:1706.06083, 2017.
[13]	MOOSAVI-DEZFOOLI S M , FAWZI A , FROSSARD P . DeepFool:a simple and accurate method to fool deep neural networks[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. 2016: 2574-2582.
[14]	SU J W , VARGAS D V , SAKURAI K . One pixel attack for fooling deep neural networks[J]. IEEE Transactions on Evolutionary Computation, 2019,23(5): 828-841.
[15]	CARLINI N , WAGNER D . Towards evaluating the robustness of neural networks[C]// Proceedings of 2017 IEEE Symposium on Security and Privacy. 2017: 39-57.
[16]	KANNAN H , KURAKIN A , GOODFELLOW I . Adversarial logit pairing[J]. arXiv preprint arXiv:1803.06373, 2018.
[17]	WANG Y , ZOU D , YI J ,et al. Improving adversarial robustness requires revisiting misclassified examples[C]// Proceedings of International Conference on Learning Representations. 2019.
[18]	XIE C , WANG J , ZHANG Z ,et al. Mitigating adversarial effects through randomization[J]. arXiv preprint arXiv:1711.01991, 2017.
[19]	XIE C H , WU Y X , VAN DER MAATEN L ,et al. Feature denoising for improving adversarial robustness[C]// Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2019: 501-509.
[20]	XU W L , EVANS D , QI Y J . Feature squeezing:detecting adversarial examples in deep neural networks[J]. arXiv preprint arXiv:1704.01155, 2017.
[21]	MA S Q , LIU Y Q , TAO G H ,et al. Nic:detecting adversarial samples with neural network invariant checking[C]// Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019). 2019.
[22]	GOLDBLUM M , FOWL L , FEIZI S ,et al. Adversarially robust distillation[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020,34(4): 3996-4003.
[23]	SHAFAHI A , SAADATPANAH P , ZHU C ,et al. Adversarially robust transfer learning[J]. arXiv preprint arXiv:1905.08232, 2019.
[24]	SHAFAHI A , NAJIBI M , GHIASI A ,et al. Adversarial training for free![J]. arXiv preprint arXiv:1904.12843, 2019.
[25]	SELVARAJU R R , COGSWELL M , DAS A ,et al. Grad-CAM:visual explanations from deep networks via gradient-based localization[J]. International Journal of Computer Vision, 2020,128(2): 336-359.
[26]	CHATTOPADHAY A , SARKAR A , HOWLADER P ,et al. Grad-CAM++:generalized gradient-based visual explanations for deep convolutional networks[C]// Proceedings of 2018 IEEE Winter Conference on Applications of Computer Vision. 2018: 839-847.
[27]	WANG Y D , ZHANG J , KAN M N ,et al. Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation[C]// Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2020: 12272-12281.
[28]	LIN M , CHEN Q , YAN S . Network in network[J]. arXiv preprint arXiv:1312.4400, 2013.
[29]	BOTTOU L , . Large-scale machine learning with stochastic gradient descent[C]// Proceedings of COMPSTAT'2010. 2010.