The normal mode interference characteristic in shallow water waveguide is a valu- able topic in the fields of underwater acoustic. A method for extracting the interference components of normal modes from broadband aco...The normal mode interference characteristic in shallow water waveguide is a valu- able topic in the fields of underwater acoustic. A method for extracting the interference components of normal modes from broadband acoustic propagation data recorded by a single hy- drophone without any prior information is present in this paper. First, a Hermitian matrix is formed by the power spectral density. Second, a singular value decomposition (SVD) is performed on the Hermitian matrix to obtain the orthonormal eigenvectors, which are proportional to the interference components of normal modes. The fundamental equations of the new extracting method are derived based on normal mode and waveguide invariant theory. And the validity of the present method is verified by the numerical simulation and experimental results. In addition, the extracted results of normal-mode interference components are intended to be used for passive ranging of broadband sources.展开更多
对抗攻击的出现对于深度神经网络(DNN)在现实场景中的大规模部署产生了巨大的威胁,尤其是在与安全相关的领域。目前已有的大多数防御方法都基于启发式假设,缺少对模型对抗鲁棒性的分析。如何提升DNN的对抗鲁棒性,并提升鲁棒性的可解释...对抗攻击的出现对于深度神经网络(DNN)在现实场景中的大规模部署产生了巨大的威胁,尤其是在与安全相关的领域。目前已有的大多数防御方法都基于启发式假设,缺少对模型对抗鲁棒性的分析。如何提升DNN的对抗鲁棒性,并提升鲁棒性的可解释性和可信度,成为人工智能安全领域的重要一环。文中提出从奇异值分布的角度分析模型的对抗鲁棒性。研究发现,模型在对抗性环境下鲁棒性的提升伴随着更加平滑的奇异值分布。通过进一步分析表明,平滑的奇异值分布意味着模型的分类置信度来源更加多样,从而也具有更高的对抗鲁棒性。基于此分析,进一步提出了基于奇异值抑制SVS(Singular Value Suppress)的对抗训练方法。实验结果表明,该方法进一步提高了模型在对抗性环境下的鲁棒性,在面对强力白盒攻击方法PGD(Project Gradient Descent)时,在CIFAR10和SVHN数据集上分别能达到55.3%和54.51%的精度,超过了目前最具有代表性的对抗训练方法。展开更多
目的针对第2代数字水印技术,提出一种基于Harris特征点和DWT-SVD的图像盲水印算法。方法提取归一化图像的Harris特征点;选取部分稳定特征点来确定要嵌入水印的特征区域;将特征区域作一次小波分解得到的低频子带,对低频子带进行分块,并...目的针对第2代数字水印技术,提出一种基于Harris特征点和DWT-SVD的图像盲水印算法。方法提取归一化图像的Harris特征点;选取部分稳定特征点来确定要嵌入水印的特征区域;将特征区域作一次小波分解得到的低频子带,对低频子带进行分块,并对每一块进行奇异值分解,通过对每块中最大奇异值进行加权的方法来嵌入水印信息。结果 PSNR值均大于45 d B,NC值接近于1,说明该算法具有可行性。结论该算法对剪切攻击具有很好的鲁棒性,同时该算法也能很好地抵抗噪声、中值滤波攻击、提高亮度攻击、降低亮度攻击、基本图像处理操作的攻击。展开更多
文摘The normal mode interference characteristic in shallow water waveguide is a valu- able topic in the fields of underwater acoustic. A method for extracting the interference components of normal modes from broadband acoustic propagation data recorded by a single hy- drophone without any prior information is present in this paper. First, a Hermitian matrix is formed by the power spectral density. Second, a singular value decomposition (SVD) is performed on the Hermitian matrix to obtain the orthonormal eigenvectors, which are proportional to the interference components of normal modes. The fundamental equations of the new extracting method are derived based on normal mode and waveguide invariant theory. And the validity of the present method is verified by the numerical simulation and experimental results. In addition, the extracted results of normal-mode interference components are intended to be used for passive ranging of broadband sources.
文摘对抗攻击的出现对于深度神经网络(DNN)在现实场景中的大规模部署产生了巨大的威胁,尤其是在与安全相关的领域。目前已有的大多数防御方法都基于启发式假设,缺少对模型对抗鲁棒性的分析。如何提升DNN的对抗鲁棒性,并提升鲁棒性的可解释性和可信度,成为人工智能安全领域的重要一环。文中提出从奇异值分布的角度分析模型的对抗鲁棒性。研究发现,模型在对抗性环境下鲁棒性的提升伴随着更加平滑的奇异值分布。通过进一步分析表明,平滑的奇异值分布意味着模型的分类置信度来源更加多样,从而也具有更高的对抗鲁棒性。基于此分析,进一步提出了基于奇异值抑制SVS(Singular Value Suppress)的对抗训练方法。实验结果表明,该方法进一步提高了模型在对抗性环境下的鲁棒性,在面对强力白盒攻击方法PGD(Project Gradient Descent)时,在CIFAR10和SVHN数据集上分别能达到55.3%和54.51%的精度,超过了目前最具有代表性的对抗训练方法。
文摘目的针对第2代数字水印技术,提出一种基于Harris特征点和DWT-SVD的图像盲水印算法。方法提取归一化图像的Harris特征点;选取部分稳定特征点来确定要嵌入水印的特征区域;将特征区域作一次小波分解得到的低频子带,对低频子带进行分块,并对每一块进行奇异值分解,通过对每块中最大奇异值进行加权的方法来嵌入水印信息。结果 PSNR值均大于45 d B,NC值接近于1,说明该算法具有可行性。结论该算法对剪切攻击具有很好的鲁棒性,同时该算法也能很好地抵抗噪声、中值滤波攻击、提高亮度攻击、降低亮度攻击、基本图像处理操作的攻击。