受对抗样本自身可迁移属性的影响,传统对抗样本防御方法的防御效果存在不稳定的情况,为此,提出基于深度学习的对抗样本防御方法。文章借助深度学习算法,构建了对抗样本伊辛模型,设置模型的初始状态为神经网络的输入数据,采用自旋状态表...受对抗样本自身可迁移属性的影响,传统对抗样本防御方法的防御效果存在不稳定的情况,为此,提出基于深度学习的对抗样本防御方法。文章借助深度学习算法,构建了对抗样本伊辛模型,设置模型的初始状态为神经网络的输入数据,采用自旋状态表示每一个神经元值与对抗样本伊辛模型的格点,并利用神经网络中卷积运算的特征,消解势场中预先给定的外部磁化作用,以最大限减少低对抗样本伊辛模型在能量作用下的局部自旋问题。在对抗样本防御阶段,利用对抗样本伊辛模型的通道相关性,生成重要性掩码对通道的激活进行调整,并结合对抗样本伊辛模型通道梯度累积值的实际情况设置了差异化的重要性掩码生成函数。在应用测试过程中,为验证防御效果,在快速梯度下降法(Fast Gradient Sign Method,FGSM)、Deepfool、C&W(Carlini and Wagner)攻击算法、投影梯度下降(Projected Gradient Descent,PFD)、集成对抗检测器(Energy-Aware Data-centric,EAD)共5种对抗策略下设计了对抗样本防御方法,对比不同对抗样本防御方法的性能,发现文章提出的基于深度学习的对抗样本防御方法的曲线下的面积(Area Under the Curve,AUC)值稳定在0.95以上,说明对抗样本防御方法具有较好的防御性能。展开更多
In this paper, a close-loop feedback control is imposed locally on the Fitzhugh-Nagumo （FHN） system to suppress the stable spirals and spatiotemporal chaos according to the principle of self-adaptive coupling intera...In this paper, a close-loop feedback control is imposed locally on the Fitzhugh-Nagumo （FHN） system to suppress the stable spirals and spatiotemporal chaos according to the principle of self-adaptive coupling interaction. The simulation results show that an expanding target wave is stimulated by the spiral waves under dynamic control period when a local area. of 5 × 5 grids is controlled, or the spiral tip is driven to the board of the system, It is also found that the spatiotemporal chaos can be suppressed to get a stable homogeneous state within 50 time units as two local grids are controlled mutually. The mechanism of the scheme is briefly discussed.展开更多
We explore the spin-boson model in a special case, i.e., with zero local field. In contrast to previous studies, we find no possibility for quantum phase transition （QPT） happening between the localized and delocali...We explore the spin-boson model in a special case, i.e., with zero local field. In contrast to previous studies, we find no possibility for quantum phase transition （QPT） happening between the localized and delocalized phases, and the behavior of the model can be fully characterized by the even or odd parity as well as the parity breaking, instead of the QPT, owned by the ground state of the system. The parity breaking mentioned in our case is completely different from the spontaneously broken symmetry relevant to the conventionally defined QPT in previous studies. Our analytical treatment about the eigensolution of the ground state of the model presents for the first time a rigorous proof of no- degeneracy for the ground state of the model, which is independent of the bath type, the degrees of freedom of the bath and the calculation precision. We argue that the QPT mentioned previously appears due to incorrect employment of the ground state of the model and/or unreasonable treatment of the infrared divergence existing in the spectral functions for Ohmic and sub-Ohmic dissipations.展开更多
文摘受对抗样本自身可迁移属性的影响,传统对抗样本防御方法的防御效果存在不稳定的情况,为此,提出基于深度学习的对抗样本防御方法。文章借助深度学习算法,构建了对抗样本伊辛模型,设置模型的初始状态为神经网络的输入数据,采用自旋状态表示每一个神经元值与对抗样本伊辛模型的格点,并利用神经网络中卷积运算的特征,消解势场中预先给定的外部磁化作用,以最大限减少低对抗样本伊辛模型在能量作用下的局部自旋问题。在对抗样本防御阶段,利用对抗样本伊辛模型的通道相关性,生成重要性掩码对通道的激活进行调整,并结合对抗样本伊辛模型通道梯度累积值的实际情况设置了差异化的重要性掩码生成函数。在应用测试过程中,为验证防御效果,在快速梯度下降法(Fast Gradient Sign Method,FGSM)、Deepfool、C&W(Carlini and Wagner)攻击算法、投影梯度下降(Projected Gradient Descent,PFD)、集成对抗检测器(Energy-Aware Data-centric,EAD)共5种对抗策略下设计了对抗样本防御方法,对比不同对抗样本防御方法的性能,发现文章提出的基于深度学习的对抗样本防御方法的曲线下的面积(Area Under the Curve,AUC)值稳定在0.95以上,说明对抗样本防御方法具有较好的防御性能。
基金The project supported partially by National Natural Science Foundation of China under Grant No. 90303010 We would like to thank H.Zhang for valuable discussions.
文摘In this paper, a close-loop feedback control is imposed locally on the Fitzhugh-Nagumo （FHN） system to suppress the stable spirals and spatiotemporal chaos according to the principle of self-adaptive coupling interaction. The simulation results show that an expanding target wave is stimulated by the spiral waves under dynamic control period when a local area. of 5 × 5 grids is controlled, or the spiral tip is driven to the board of the system, It is also found that the spatiotemporal chaos can be suppressed to get a stable homogeneous state within 50 time units as two local grids are controlled mutually. The mechanism of the scheme is briefly discussed.
基金Supported by National Fundamental Research Program of China under Grant No.2012CB922102National Natural Science Foundation of China under Grant Nos.10974225 and 11004226State Key Laboratory Funding of WIPM
文摘We explore the spin-boson model in a special case, i.e., with zero local field. In contrast to previous studies, we find no possibility for quantum phase transition （QPT） happening between the localized and delocalized phases, and the behavior of the model can be fully characterized by the even or odd parity as well as the parity breaking, instead of the QPT, owned by the ground state of the system. The parity breaking mentioned in our case is completely different from the spontaneously broken symmetry relevant to the conventionally defined QPT in previous studies. Our analytical treatment about the eigensolution of the ground state of the model presents for the first time a rigorous proof of no- degeneracy for the ground state of the model, which is independent of the bath type, the degrees of freedom of the bath and the calculation precision. We argue that the QPT mentioned previously appears due to incorrect employment of the ground state of the model and/or unreasonable treatment of the infrared divergence existing in the spectral functions for Ohmic and sub-Ohmic dissipations.
