人工智能培训

搜索

机器学习论文:基于鲁棒张量网络分解的变换输入卷积神经网络(Convolutional Neural Networks with Transformed Inpu

[复制链接]
1yuqunchenggong 发表于 2018-12-7 11:36:01 | 显示全部楼层 |阅读模式
1yuqunchenggong 2018-12-7 11:36:01 31 0 显示全部楼层
机器学习论文:基于鲁棒张量网络分解的变换输入卷积神经网络(Convolutional Neural Networks with Transformed Input based on Robust  Tensor Network Decomposition)张量网络分解,源于量子物理学到模式化的多粒子量子系统,被证明是一种有前途的数学技术,能够以简洁的方式有效地表示和处理大数据。在这项研究中,我们表明张量网络可以系统地划分结构化数据,例如彩色图像,用于分散存储和以隐私保护方式进行通信。利用大海数据和元数据隐私,实证结果表明,存在于张量网络格式中的隐含信息的邻居扩展器无法识别数据重建。该技术补充了现有的加密技术和随机化技术,这些技术在一个地方存储明确的数据表示,并且极易受到对抗性攻击和去匿名化的攻击。此外,我们提出了一种对抗性实例的理论,该理论误导了基于奇异值分解(SVD)的子空间分析的卷积神经网络tomisclassification。该理论扩展到使用传动系统SVD(TT-SVD)分析高阶张量;它有助于解释不同数据集对敌对攻击的敏感程度,包括全局和局部攻击在内的不同对抗性攻击的结构相似性,以及基于输入变换的不同对抗性防御的效果。然后开发了基于鲁棒TT-SVD的高效自适应算法来检测强对抗和静态对抗攻击。
Tensor network decomposition, originated from quantum physics to modelentangled many-particle quantum systems, turns out to be a promisingmathematical technique to efficiently represent and process big data inparsimonious manner.In this study, we show that tensor networks cansystematically partition structured data, e.g.color images, for distributedstorage and communication in privacy-preserving manner.Leveraging the sea ofbig data and metadata privacy, empirical results show that neighbouringsubtensors with implicit information stored in tensor network formats cannot beidentified for data reconstruction.This technique complements the existingencryption and randomization techniques which store explicit datarepresentation at one place and highly susceptible to adversarial attacks suchas side-channel attacks and de-anonymization.Furthermore, we propose a theoryfor adversarial examples that mislead convolutional neural networks tomisclassification using subspace analysis based on singular value decomposition(SVD).The theory is extended to analyze higher-order tensors usingtensor-train SVD (TT-SVD);it helps to explain the level of susceptibility ofdifferent datasets to adversarial attacks, the structural similarity ofdifferent adversarial attacks including global and localized attacks, and theefficacy of different adversarial defenses based on input transformation.Anefficient and adaptive algorithm based on robust TT-SVD is then developed todetect strong and static adversarial attacks.机器学习论文:基于鲁棒张量网络分解的变换输入卷积神经网络(Convolutional Neural Networks with Transformed Input based on Robust  Tensor Network Decomposition) ZLC45cN6D4ws6cC5.jpg
URL地址:https://arxiv.org/abs/1812.02622     ----pdf下载地址:https://arxiv.org/pdf/1812.02622    ----机器学习论文:基于鲁棒张量网络分解的变换输入卷积神经网络(Convolutional Neural Networks with Transformed Input based on Robust  Tensor Network Decomposition)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

1yuqunchenggong当前离线
新手上路

查看:31 | 回复:0

快速回复 返回顶部 返回列表