深度神经网络具有脆弱性,容易被精心设计的对抗样本攻击.梯度攻击方法在白盒模型上攻击成功率较高,但在黑盒模型上的迁移性较弱.基于Heavy-ball型动量和Nesterov型动量的梯度攻击方法由于在更新方向上考虑了历史梯度信息,提升了对抗样...深度神经网络具有脆弱性,容易被精心设计的对抗样本攻击.梯度攻击方法在白盒模型上攻击成功率较高,但在黑盒模型上的迁移性较弱.基于Heavy-ball型动量和Nesterov型动量的梯度攻击方法由于在更新方向上考虑了历史梯度信息,提升了对抗样本的迁移性.为了进一步使用历史梯度信息,本文针对收敛性更好的Nesterov型动量方法,使用自适应步长策略代替目前广泛使用的固定步长,提出了一种方向和步长均使用历史梯度信息的迭代快速梯度方法(Nesterov and Adaptive-learning-rate based Iterative Fast Gradient Method,NAI-FGM).此外,本文还提出了一种线性变换不变性(Linear-transformation Invariant Method,LIM)的数据增强方法 .实验结果证实了NAI-FGM攻击方法和LIM数据增强策略相对于同类型方法均具有更高的黑盒攻击成功率.组合NAI-FGM方法和LIM策略生成对抗样本,在常规训练模型上的平均黑盒攻击成功率达到87.8%,在对抗训练模型上的平均黑盒攻击成功率达到57.5%,在防御模型上的平均黑盒攻击成功率达到67.2%,均超过现有最高水平.展开更多
Hot nuclei produced in the reactions of 25 MeV/u<sup>40</sup>Ar with <sup>209</sup>Bi and<sup>Na1</sup>Ag are studied.Theinitial properties of these nuclei :excitation energies,line...Hot nuclei produced in the reactions of 25 MeV/u<sup>40</sup>Ar with <sup>209</sup>Bi and<sup>Na1</sup>Ag are studied.Theinitial properties of these nuclei :excitation energies,linear momentum transfer and temperatures are char-acterized through the measurements of folding angle of fission fragments,light charged particle energyspectra and intermediate mass fragments.Nuclei with excitation energies as high as～600 MeV and tem-peratures as high as～6 MeV are produced.The spectra of alpha particles detected in coincidence with fis-sion fragments in the case of different average linear momentum transfer are analyzed and the initial tem-peratures of the hot nuclei are determined.A saturation（or limiting）temperature near 6 MeV is foundfor alpha emission from medium mass nuclei having excitation energies higher than 3 MeV/u.This resultprovides evidence for a soft nuclear equation of state at thee high excitation energies and is consistentwith predictions of statistic multifragmentation model calculations.展开更多
文摘深度神经网络具有脆弱性,容易被精心设计的对抗样本攻击.梯度攻击方法在白盒模型上攻击成功率较高,但在黑盒模型上的迁移性较弱.基于Heavy-ball型动量和Nesterov型动量的梯度攻击方法由于在更新方向上考虑了历史梯度信息,提升了对抗样本的迁移性.为了进一步使用历史梯度信息,本文针对收敛性更好的Nesterov型动量方法,使用自适应步长策略代替目前广泛使用的固定步长,提出了一种方向和步长均使用历史梯度信息的迭代快速梯度方法(Nesterov and Adaptive-learning-rate based Iterative Fast Gradient Method,NAI-FGM).此外,本文还提出了一种线性变换不变性(Linear-transformation Invariant Method,LIM)的数据增强方法 .实验结果证实了NAI-FGM攻击方法和LIM数据增强策略相对于同类型方法均具有更高的黑盒攻击成功率.组合NAI-FGM方法和LIM策略生成对抗样本,在常规训练模型上的平均黑盒攻击成功率达到87.8%,在对抗训练模型上的平均黑盒攻击成功率达到57.5%,在防御模型上的平均黑盒攻击成功率达到67.2%,均超过现有最高水平.
基金The project supported by the National Natural Science Foundation of China
文摘Hot nuclei produced in the reactions of 25 MeV/u<sup>40</sup>Ar with <sup>209</sup>Bi and<sup>Na1</sup>Ag are studied.Theinitial properties of these nuclei :excitation energies,linear momentum transfer and temperatures are char-acterized through the measurements of folding angle of fission fragments,light charged particle energyspectra and intermediate mass fragments.Nuclei with excitation energies as high as～600 MeV and tem-peratures as high as～6 MeV are produced.The spectra of alpha particles detected in coincidence with fis-sion fragments in the case of different average linear momentum transfer are analyzed and the initial tem-peratures of the hot nuclei are determined.A saturation（or limiting）temperature near 6 MeV is foundfor alpha emission from medium mass nuclei having excitation energies higher than 3 MeV/u.This resultprovides evidence for a soft nuclear equation of state at thee high excitation energies and is consistentwith predictions of statistic multifragmentation model calculations.
