Mixture ratio modeling of dynamic systems

Any knowledge extraction relies (possibly implicitly) on a hypothesis about the modelled-data dependence. The extracted knowledge ultimately serves to a decision-making (DM). DM always faces uncertainty and this makes probabilistic modelling adequate. The inspected black-box modeling deals with “universal” approximators of the relevant probabilistic model. Finite mixtures with components in the exponential family are often exploited. Their attractiveness stems from their flexibility, the cluster interpretability of components and the existence of algorithms for processing high-dimensional data streams. They are even used in dynamic cases with mutually dependent data records while regression and auto-regression mixture components serve to the dependence modeling. These dynamic models, however, mostly assume data-independent component weights, that is, memoryless transitions between dynamic mixture components. Such mixtures are not universal approximators of dynamic probabilistic models. Formally, this follows from the fact that the set of finite probabilistic mixtures is not closed with respect to the conditioning, which is the key estimation and predictive operation. The paper overcomes this drawback by using ratios of finite mixtures as universally approximating dynamic parametric models. The paper motivates them, elaborates their approximate Bayesian recursive estimation and reveals their application potential.     

NP问题的最优轮复杂性知识的零知识证明

NP问题已有的知识的(黑箱)零知识证明都是非常数轮的,因此,在标准的复杂性假设下,NP问题是否存在常数轮的(黑箱)知识的零知识证明是一个有意义的问题.本文对该问题进行了研究,在一定的假设下给出了HC问题的两个常数轮知识的零知识证明系统.根据Katz最近的研究结果,在多项式分层不坍塌的条件下,本文基于claw-free陷门置换给出的HC问题的5轮知识的零知识证明系统具有最优的轮复杂性.     

Simplex噪声区域中目标特征的对抗攻击算法

针对当前黑盒环境中,主流的图像对抗攻击算法在有限的目标模型访问查询次数条件下攻击准确率低的问题,提出一种基于目标特征和限定区域采样的目标攻击算法.首先根据原始图像和目标图像生成初始对抗样本;然后在Simplex-mean噪声区域中进行扰动采样,并根据对抗样本和原始图像差异度以及目标特征区域位置决定扰动大小;最后将扰动作用于初始对抗样本中,使新的对抗样本在保持对抗性的同时缩小与原始图像的差异度.以常见的图像分类模型InceptionV3和VGG16等为基础,在相同的目标模型访问查询,以及与对抗样本和原始图像的l₂距离小于55.89的条件下,采用BBA等算法对同一图像集和目标集进行攻击.实验结果表明,在同样的目标模型访问查询和l₂=55.89的限制条件下,不超过5 000次目标查询时,在InceptionV3模型上该算法的攻击准确率比同类攻击算法提升至少50%.     

颜色模型扰动的语义对抗样本生成方法

卷积神经网络是一种具有强大特征提取能力的深度神经网络,其在众多领域得到了广泛应用。但是,研究表明卷积神经网络易受对抗样本攻击。不同于传统的以梯度迭代生成对抗扰动的方法,提出了一种基于颜色模型的语义对抗样本生成方法,利用人类视觉和卷积模型在识别物体中表现出的形状偏好特性,通过颜色模型的扰动变换来生成对抗样本。在样本生成过程中其不需要目标模型的网络参数、损失函数或者相关结构信息,仅依靠颜色模型的变换和通道信息的随机扰动,所以这是一种可以完成黑盒攻击的对抗样本。     

基于深度神经网络的对抗样本技术综述

深度学习在完成一些难度极高的任务中展现了惊人的能力,但深度神经网络难以避免对刻意添加了扰动的样本（称为“对抗样本”）进行错误的分类。“对抗样本”逐渐成为深度学习安全领域的研究热点。研究对抗样本产生的原因和作用机理,有助于从安全性和鲁棒性方面优化模型。在掌握对抗样本原理的基础上,对经典对抗样本攻击方法进行分类总结,根据不同的攻击原理将攻击方法分为白盒攻击与黑盒攻击两个大类,并引入非特定目标攻击、特定目标攻击、全像素添加扰动攻击和部分像素添加扰动攻击等细类。在ImageNet数据集上对几种典型攻击方法进行复现,通过实验结果,比较几种生成方法的优缺点,分析对抗样本生成过程中的突出问题。并对对抗样本的应用和发展作了展望。     

嵌入式汇编语言软件单元测试方法

单元测试是在整个软件开发过程中的一个重要环节,单元是软件的构成基础,因此单元的质量是整个软件质量的保证。嵌入式软件由于与CPu密切相关,多采用汇编语言编程,使得其测试一直是个难点,特别是单元测试。针对嵌入式软件的特点,本文根据软件单元测试的理论,介绍一种使用旋极AsmTester对嵌入式汇编语言编程的软件进行单元测试的方法。     

基于群的公钥叛逆者追踪方案

为有效保护数据提供者的合法权益,提出一种基于群的公钥叛逆者追踪方案。依据群合谋的行为特征将合法用户分成不同的群,利用分群思想生成更多合法密钥。在叛逆者检测时将可疑用户缩小到特定群,从而提高检测效率。通过求解大整数分解困难问题和离散对数困难问题证明,该方案完全抗合谋,且具有黑盒追踪性。     

实现分离多级存取结构的黑盒密钥共享体制

黑盒密钥共享体制和有限域上的密钥共享体制是不同的.该体制只要求主密钥空间是一个有限交换群,并且可以将群运算和随机挑选群元素等过程作为黑盒调用.尤其是,它与群的结构和除数无关.迄今为止,还没有造出有效的实现非门限存取结构的黑盒密钥共享体制.文中给出了任意交换群上的实现一类非门限存取结构的黑盒密钥共享体制的构造方法.作者使用的工具是环上单调张成方案.首先,作者提出弱单调张成方案的概念,并给出了使用弱单调张成方案构造单调张成方案的方法.然后,利用有理数域上的单调张成方案和有限域上的单调张成方案,构造出了一对互素的弱单调张成方案,最终构造出了实现一类非门限存取结构——分离多级存取结构的有效黑盒密钥共享体制.作者构造的新体制可用于构造新的实现分离多级存取结构的环上安全多方计算协议、线性整密钥共享体制、分布式RSA签名协议和新的零知识证明协议.     

Black-box models for fault detection and performance monitoring of buildings

Black-box models such as linear regression have proven to be helpful in ongoing building commissioning in many ways. The aim of this work is to improve linear models with change point for fault detection in buildings. Building simulations revealed poor performance of them (R 2 2. Remark: In the Inverse Modelling Toolkit (IMT) (Kissock et al. 2003) similar models are called CP-MVR or rather 3P-MVR or 4P-MVR. The difference to the CPMLRs used here is that there are different intercept terms for all linear regressions for the CPMLRs instead of one equal intercept term for all linear regression models 3P-MVR. This is introducing discontinuities on continuous independent variables but for the mean daily perspective taken here this is not disadvantageous.  < 0.7) for some low energy buildings. The regression models (RMs) can be considerably improved by introducing the rate of change of the indoor air temperature (ΔT _ind) as an independent variable. Thus, R ² values were raised by up to 0.5 (e.g. from 0.2 to 0.7, example with the lowest R ²). A new training and application process for the RMs revealed further improvements by using a hierarchical agglomerative clustering algorithm to determine different day-types as additional (categorical) variables in the RM. The application of these improved RMs for outlier detection is demonstrated in three buildings.     

Multilevel surrogate modeling of an amine scrubbing process for CO2capture

Surrogate models provide a powerful method for simplifying calculations within complex simulations. While surrogate models are broadly applied within chemical engineering, little research exists investigating the level of surrogacy's impact on a simplified process model. In this work, artificial neural networks (ANN) and Kriging models are used as surrogate models at the process, process unit, and thermodynamic levels for a CO₂ amine scrubbing process. The surrogated models are evaluated against an Aspen Plus simulation for accuracy, convergence behavior, computational cost, and ability to extrapolate. The thermodynamic and process unit models can better handle discontinuous, non-smooth behavior, and convergence issues in the surrogated truth model, but poor conditioning in the final system of equations results in a lower accuracy and convergence rate than the process level surrogate. Beyond model accuracy, availability of diverse data, intended re-usability, and the desired outputs must be considered when selecting a level of abstraction.