Recently published results of field and laboratory experiments on the seismic/acoustic response to injection of direct current （DC） pulses into the Earth crust or stressed rock samples raised a question on a possibi...Recently published results of field and laboratory experiments on the seismic/acoustic response to injection of direct current （DC） pulses into the Earth crust or stressed rock samples raised a question on a possibility of electrical earthquake triggering. A physical mechanism of the considered phenomenon is not clear yet in view of the very low current density （10-7-10-s A/m^2） generated by the pulsed power systems at the epicenter depth （5-10 km） of local earthquakes occurred just after the current injection. The paper describes results of laboratory ＂earthquake＂ triggering by DC pulses under conditions of a spring-block model simulated the seismogenic fault. It is experimentally shown that the electric triggering of the laboratory ＂earthquake＂ （sharp slip of a movable block of the spring-block system） is possible only within a range of subcritical state of the system, when the shear stress between the movable and fixed blocks obtains 0.98-0.99 of its critical value. The threshold of electric triggering action is about 20 A/m^2 that is 7-8 orders of magnitude higher than estimated electric current density for Bishkek test site （Northern Tien Shan, Kirghizia） where the seismic response to the man-made electric action was observed. In this connection, the electric triggering phenomena may be explained by contraction of electric current in the narrow conductive areas of the faults and the corresponding increase in current density or by involving the secondary triggering mechanisms like electromagnetic stimulation of conductive fluid migration into the fault area resulted in decrease in the fault strength properties.展开更多
后门攻击是机器学习模型在训练阶段面临的主要安全威胁之一。尽管现有针对后门攻击的防御研究已取得显著进展,但这些方法往往会导致模型在干净测试集上的准确率出现明显下降。为解决这一问题,提出了一种基于扩散模型的机器学习后门攻击...后门攻击是机器学习模型在训练阶段面临的主要安全威胁之一。尽管现有针对后门攻击的防御研究已取得显著进展,但这些方法往往会导致模型在干净测试集上的准确率出现明显下降。为解决这一问题,提出了一种基于扩散模型的机器学习后门攻击防御方法(defending against backdoor attacks with diffusion model,DBADM)。该方法的核心思想是在模型训练前,利用扩散模型对含有后门触发器的中毒样本进行预处理,通过改变样本中隐藏的触发器特征,从而有效抵御后门攻击。研究人员在MNIST、CIFAR-10、Tiny ImageNet和LFW4个基准数据集上进行了系统的攻防对比实验。实验结果表明,所提出的DBADM方法不仅能够有效防御各类后门攻击,还能保持模型在干净数据集上的高精度性能。展开更多
基金funded by Russian Foundation for Basic Research according to research project No.15-55-53104National Natural Science Foundation of China according to International cooperation project No.41511130032
文摘Recently published results of field and laboratory experiments on the seismic/acoustic response to injection of direct current （DC） pulses into the Earth crust or stressed rock samples raised a question on a possibility of electrical earthquake triggering. A physical mechanism of the considered phenomenon is not clear yet in view of the very low current density （10-7-10-s A/m^2） generated by the pulsed power systems at the epicenter depth （5-10 km） of local earthquakes occurred just after the current injection. The paper describes results of laboratory ＂earthquake＂ triggering by DC pulses under conditions of a spring-block model simulated the seismogenic fault. It is experimentally shown that the electric triggering of the laboratory ＂earthquake＂ （sharp slip of a movable block of the spring-block system） is possible only within a range of subcritical state of the system, when the shear stress between the movable and fixed blocks obtains 0.98-0.99 of its critical value. The threshold of electric triggering action is about 20 A/m^2 that is 7-8 orders of magnitude higher than estimated electric current density for Bishkek test site （Northern Tien Shan, Kirghizia） where the seismic response to the man-made electric action was observed. In this connection, the electric triggering phenomena may be explained by contraction of electric current in the narrow conductive areas of the faults and the corresponding increase in current density or by involving the secondary triggering mechanisms like electromagnetic stimulation of conductive fluid migration into the fault area resulted in decrease in the fault strength properties.
文摘后门攻击是机器学习模型在训练阶段面临的主要安全威胁之一。尽管现有针对后门攻击的防御研究已取得显著进展,但这些方法往往会导致模型在干净测试集上的准确率出现明显下降。为解决这一问题,提出了一种基于扩散模型的机器学习后门攻击防御方法(defending against backdoor attacks with diffusion model,DBADM)。该方法的核心思想是在模型训练前,利用扩散模型对含有后门触发器的中毒样本进行预处理,通过改变样本中隐藏的触发器特征,从而有效抵御后门攻击。研究人员在MNIST、CIFAR-10、Tiny ImageNet和LFW4个基准数据集上进行了系统的攻防对比实验。实验结果表明,所提出的DBADM方法不仅能够有效防御各类后门攻击,还能保持模型在干净数据集上的高精度性能。
