Black-box Features Research Articles

Improving the transferability of adversarial examples through black-box feature attacks

Deep neural networks (DNNs) are vulnerable and susceptible to imperceptible perturbations. Adversarial examples become more and more popular. Black-box attacks are considered to be the most realistic scenario. Currently, transfer-based black-box attacks show excellent performance. However, transfer-based black-box attacks all require an agent model of the attack, which we call the source model. This leads to the existing transfer-based attacks limited by the features focused on the source model, which creates a bottleneck in improving the transferability of adversarial examples. In order to solve this problem, we propose an attack that mainly targets features that are insensitive to the source model, which we call the black-box feature attack. Specifically, we categorize the features of the image into white-box features and black-box features. The white-box features are source model-sensitive features and the black-box features are source model insensitive features. White-box features are only specific to the source model, while black-box features are more generalized for unknown models. By destroying the image white-box features, the fitted image is obtained and the model intermediate layer feature map is extracted. Afterward, the fitting gradient is found for the fitted images with different fitting degrees. We construct loss functions based on the obtained fitting gradients and feature maps to guide the attacks to better destroy the black-box features of the images. Extensive experiments demonstrate that our methods have higher transferability compared to state-of-the-art methods, which achieve more than 90% of transferability under the normal model. It is also significantly better than other methods on adversarially trained models. Even in the white-box setting, our attack has the best performance.

Neural Network Design for Impedance Modeling of Power Electronic Systems Based on Latent Features.

Data-driven approaches are promising to address the modeling issues of modern power electronics-based power systems, due to the black-box feature. Frequency-domain analysis has been applied to address the emerging small-signal oscillation issues caused by converter control interactions. However, the frequency-domain model of a power electronic system is linearized around a specific operating condition. It thus requires measurement or identification of frequency-domain models repeatedly at many operating points (OPs) due to the wide operation range of the power systems, which brings significant computation and data burden. This article addresses this challenge by developing a deep learning approach using multilayer feedforward neural networks (FNNs) to train the frequency-domain impedance model of power electronic systems that is continuous of OP. Distinguished from the prior neural network designs relying on trial-and-error and sufficient data size, this article proposes to design the FNN based on latent features of power electronic systems, i.e., the number of system poles and zeros. To further investigate the impacts of data quantity and quality, learning procedures from a small dataset are developed, and K-medoids clustering based on dynamic time warping is used to reveal insights into multivariable sensitivity, which helps improve the data quality. The proposed approaches for the FNN design and learning have been proven simple, effective, and optimal based on case studies on a power electronic converter, and future prospects in its industrial applications are also discussed.

PPS: Fair and efficient black-box scheduling for multi-tenant GPU clusters

Multi-tenant GPU clusters are common, where users purchase GPU quota to run their neural network training jobs. However, strict quota-based scheduling often leads to cluster under-utilization, while allowing quota groups to use excess GPUs improves utilization but results in fairness problems. We propose PPS, a probabilistic prediction based scheduler, which uses job history statistics to predict future cluster status for making good scheduling decisions. Different from existing schedulers that rely on deep learning frameworks to adjust bad scheduling decisions and/or require detailed job information, PPS treats jobs as black boxes in that PPS runs a job to completion without adjustment once scheduled and requires only aggregate job statistics. The black-box feature is favorable due to its good generality, compatibility and security, and made possible by the predictability of aggregate resource utilization statistics of large clusters. Extensive experiments show that PPS achieves high cluster utilization and good fairness simultaneously.

Comparison of Transfer Function Models to Represent the Correlation Between Vehicle Lateral Acceleration and Head Tilting Angle in Motion Sickness

Motion Sickness (MS) is described as an unpleasant feeling caused by a forceful movement; hence vehicle movement impacts the severity of MS. While negotiating a curve, drivers and passenger tilt their heads differently, affecting their motion sickness incidence (MSI), which is the severity of MS. MS is a negative feeling, that affects occupant’s comfort, and to further understand the correlation between occupants' behavior and vehicle movement in MS and then represent it using mathematical models, it was proven that MSI could be predicted through mathematical models. However, there is an indefinite value between values between occupant’s behavior and vehicle movement. Based on that it is vital to express it the correlation mathematically. An experiment adopted from a prior study was utilized to get the data and develope the mathematical models with different proportions to represent the correlation between vehicle movement and occupant behavior in motion sickness in transfer function equations using system identification (SI), by utilising black-box feature to use the experimental data as input and output to allow SI to predict the transfer function models. The aim of this study is to investigate MS factors in relation to the vehicle movement and occupant’s behavior, to develop multiple transfer function models, to analyze and compare them. The results were obtained in the three different transfer function orders, second, third and fourth order functions for each proportion used for both the driver and passenger, the driver models’ results were in between 64.68%-67.87%, and the passenger results were in between 63.75%-67.93%, after the comparison the highest fits for each order were obtained. The highest fits amongst driver models were 67.87% (4th Order), 66.78% (3rd Order) and 65.17% (2nd Order) and 67.93% (4th Order), 66.3% (3rd Order) and 64.82% (2nd Order) amongst the passenger models. Those fits were then validated via Simulink with unseen data that was not used in identification process, and lastly the models Root Mean Square Error (RMSE) was obtained for all of them to determine their efficiency.

Safety Performance Boundary Identification of Highly Automated Vehicles: A Surrogate Model-Based Gradient Descent Searching Approach

Highly automated vehicles (HAVs) have been introduced to the transportation system for the purpose of providing safer mobility. Considering the expected long co-existence period of HAVs and human-driven vehicles (HDVs), the safety operation of HAVs interacting with HDVs needs to be verified. To achieve this, HAVs’ Operational Design Domain (ODD) needs to be identified under the scenario-based testing framework. In this study, a novel testing framework aiming at identifying the Safety performance boundary (SPB) is proposed, which assures the coverage of safety-critical scenarios and compatible with the black-box feature of HAV control algorithm. A surrogate model was utilized to approximate the safety performance of HAV, and a gradient descent searching algorithm was employed to accelerate the search for SPB. For empirical analyses, a three-vehicle following scenario was adopted and the Intelligent Driver Model (IDM) was tested as a case study. The results show that only 4% of the total scenarios are required to establish a reliable surrogate model. And the gradient descent algorithm was able to establish the SPB by identifying 97.42% of collision scenarios and only false alarming 0.29% of non-collision scenarios. Furthermore, the concept of safety tolerance was proposed to measure the possibilities of boundary scenarios dropping in safety performance. The applications of helping to construct ODD and compare different control algorithms were discussed. It shows that the IDM performs better than the Wiedemann 99 (W99) model with larger ODD.

Vehicle Black Box Based on Can and IOT Protocol 2

Abstract: Present Automobiles are being developed with more of electrical and electronic parts for efficient operation. Recently, a vehicle was built with an analog driver-vehicle interface for indicating various vehicle status like speed, fuel level, Engine temperature etc. This project presents the development and implementation of a digital driving system for a semi-autonomous vehicle to improve the driver-vehicle interface along with black box features. It uses an Arduino based data acquisition system that uses ADC to bring all control data from analog to digital format and visualize through LCD. The communication module used in this project is embedded networking by CAN which has efficient data transfer. It also takes feedback of vehicle conditions like Vehicle speed, Engine temperature etc. And is controlled by the main controller. Additionally, this unit is equipped with GPS, GSM, ESP01 for various other purposes. Keywords: Black Box, Passengers, Incident Detection, Data Parameters, Micro Controller, etc.

Fatigue damage detection of aerospace-grade aluminum alloys using feature-based and feature-less deep neural networks

Fatigue damage is one of the most common causes of failure in aerospace structural components. While numerical modeling and laboratory-scale experimentation provide much insight to the physics of failure evolution, it is extremely challenging to account for all variabilies that a component may be subjected to during on-field operation. Human-supervised continuous monitoring of such components using sensors, therefore, provides a much-needed alternative for reliable operation of these components. By leveraging the concepts in deep learning, such human supervision can be assisted with an automated pre-trained deep network for damage detection. To that end, this article studies two distinct deep neural network (DNN) architectures for fatigue damage detection in aluminum using ultrasonic time-series data obtained from a novel customized fatigue testing apparatus. The first DNN, called as a feature-based network, is built by using two predefined features viz. frequency domain amplitude and autocorrelation from the ultrasonic data as the inputs. The second DNN, called as a feature-less network, uses the ultrasonic data as-is without any pre-processing and relies on the black-box features generated during training. The capability of fatigue crack detection for both DNN architectures is evaluated at two distinct stages of fatigue failure. The feature-less network is observed to outperform the feature-based network with an accuracy of 94.26% and 98.94% for the two stages. The result indicates that feature-less DNNs, owing to their construction, can formulate better, albeit black-box features, and simplify the process of choosing customized signal processing methods for similar problems.

Near-Exact CCSDT Energetics from Rank-Reduced Formalism Supplemented by Non-iterative Corrections

We introduce a non-iterative energy correction, added on top of the rank-reduced coupled-cluster method with single, double, and triple substitutions, that accounts for excitations excluded from the parent triple excitation subspace. The formula for the correction is derived by employing the coupled-cluster Lagrangian formalism, with an additional assumption that the parent excitation subspace is closed under the action of the Fock operator. Owing to the rank-reduced form of the triple excitation amplitudes tensor, the computational cost of evaluating the correction scales as N7, where N is the system size. The accuracy and computational efficiency of the proposed method is assessed for both total and relative correlation energies. We show that the non-iterative correction can fulfill two separate roles. If the accuracy level of a fraction of kJ/mol is sufficient for a given system, the correction significantly reduces the dimension of the parent triple excitation subspace needed in the iterative part of the calculations. Simultaneously, it enables reproducing the exact CCSDT results to an accuracy level below 0.1 kJ/mol, with a larger, yet still reasonable, dimension of the parent excitation subspace. This typically can be achieved at a computational cost only several times larger than required for the CCSD(T) method. The proposed method retains the black-box features of the single-reference coupled-cluster theory; the dimension of the parent excitation subspace remains the only additional parameter that has to be specified.

Artificial Intelligence-based Analytics for Diagnosis of Small Bowel Enteropathies and Black Box Feature Detection.

Striking histopathological overlap between distinct but related conditions poses a disease diagnostic challenge. There is a major clinical need to develop computational methods enabling clinicians to translate heterogeneous biomedical images into accurate and quantitative diagnostics. This need is particularly salient with small bowel enteropathies; environmental enteropathy (EE) and celiac disease (CD). We built upon our preliminary analysis by developing an artificial intelligence (AI)-based image analysis platform utilizing deep learning convolutional neural networks (CNNs) for these enteropathies. Data for the secondary analysis was obtained from three primary studies at different sites. The image analysis platform for EE and CD was developed using CNNs including one with multizoom architecture. Gradient-weighted class activation mappings (Grad-CAMs) were used to visualize the models' decision-making process for classifying each disease. A team of medical experts simultaneously reviewed the stain color normalized images done for bias reduction and Grad-CAMs to confirm structural preservation and biomedical relevance, respectively. Four hundred and sixty-one high-resolution biopsy images from 150 children were acquired. Median age (interquartile range) was 37.5 (19.0-121.5) months with a roughly equal sex distribution; 77 males (51.3%). ResNet50 and shallow CNN demonstrated 98% and 96% case-detection accuracy, respectively, which increased to 98.3% with an ensemble. Grad-CAMs demonstrated models' ability to learn different microscopic morphological features for EE, CD, and controls. Our AI-based image analysis platform demonstrated high classification accuracy for small bowel enteropathies which was capable of identifying biologically relevant microscopic features and emulating human pathologist decision-making process. Grad-CAMs illuminated the otherwise "black box" of deep learning in medicine, allowing for increased physician confidence in adopting these new technologies in clinical practice.

Convolutional Autoencoder for Feature Extraction in Tactile Sensing

A common approach in the field of tactile robotics is the development of a new perception algorithm for each new application of existing hardware solutions. In this letter, we present a method of dimensionality reduction of an optical-based tactile sensor image output using a convolutional neural network encoder structure. Instead of using various complex perception algorithms, and/or manually choosing task-specific data features, this unsupervised feature extraction method allows simultaneous online deployment of multiple simple perception algorithms on a common set of black-box features. The method is validated on a set of benchmarking use cases. Contact object shape, edge position, orientation, and indentation depth are estimated using shallow neural networks and machine learning models. Furthermore, a contact force estimator is trained, affirming that the extracted features contain sufficient information on both spatial and mechanical characteristics of the manipulated object.

Hierarchical multi-view aggregation network for sensor-based human activity recognition.

Sensor-based human activity recognition aims at detecting various physical activities performed by people with ubiquitous sensors. Different from existing deep learning-based method which mainly extracting black-box features from the raw sensor data, we propose a hierarchical multi-view aggregation network based on multi-view feature spaces. Specifically, we first construct various views of feature spaces for each individual sensor in terms of white-box features and black-box features. Then our model learns a unified representation for multi-view features by aggregating views in a hierarchical context from the aspect of feature level, position level and modality level. We design three aggregation modules corresponding to each level aggregation respectively. Based on the idea of non-local operation and attention, our fusion method is able to capture the correlation between features and leverage the relationship across different sensor position and modality. We comprehensively evaluate our method on 12 human activity benchmark datasets and the resulting accuracy outperforms the state-of-the-art approaches.

A Hierarchical Oil Tank Detector With Deep Surrounding Features for High-Resolution Optical Satellite Imagery

Automatic oil tank detection plays a very important role for remote sensing image processing. To accomplish the task, a hierarchical oil tank detector with deep surrounding features is proposed in this paper. The surrounding features extracted by the deep learning model aim at making the oil tanks more easily to recognize, since the appearance of oil tanks is a circle and this information is not enough to separate targets from the complex background. The proposed method is divided into three modules: 1) candidate selection; 2) feature extraction; and 3) classification. First, a modified ellipse and line segment detector (ELSD) based on gradient orientation is used to select candidates in the image. Afterward, the feature combing local and surrounding information together is extracted to represent the target. Histogram of oriented gradients (HOG) which can reliably capture the shape information is extracted to characterize the local patch. For the surrounding area, the convolutional neural network (CNN) trained in ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC2012) contest is applied as a blackbox feature extractor to extract rich surrounding feature. Then, the linear support vector machine (SVM) is utilized as the classifier to give the final output. Experimental results indicate that the proposed method is robust under different complex backgrounds and has high detection rate with low false alarm.

Design and Development of Low Cost Navigation and Security System for Indian Fisherman Using Adrino Nano Platform

<p>The fishing industry plays a major role in development of Indian economy. The recent attacks on fishermen taking place in Indo-Srilanka and Indo-Pakistan maritime boundaries have been major concerns. These attacks are primarily caused by the lack of navigation and security features during the voyage. Hence the current situation demands the implementation of precise facilities for reducing man and material loss. This project involves the design and implementation of a Low cost Navigation and Security System for Indian fishermen on Arduino Nano platform. The system developed solves the above said issues by continously tracking the location of fishing vessel and providing minimal security features. The system ensures that navigation is in safe zone within the nation’s maritime boundary and also prevents crossover. This is acheived using GPS receiver which directly links to GPS satellites for current location of the vessel. The required data fields like the latitude and longitude data along with the time stamps are extracted from the GPS samples and used for comparision for determining the exact location of the vessel. This procedure will help in detection of corner cases when the vessel is nearing or about to crossover the maritime boundary, which cannot be marked physically. It is useful for triggering conditions like enabling or disabling fuel injection system, the warning beeps and display notifications to the fishermen. Manual override facility for restarting the engine in case of crossover for limited duration is provided. The security features like authentication for the genuine operator to get access to the engine panel, the support for distress message and the storage of the exact time stamps and GPS locations after encryption in case of initiation of transmitting distress message is provided as a blackbox feature. The passcode based mechanism allows for maximum of three attempts to unlock access to control panel. The GSM modem allows for transmission of distress message to the registered base station/coast guard. The encrypted GPS samples and time stamps are stored in on-chip EEPROM memory for future reference. <strong></strong></p>

基于NARX模型的光伏并网逆变器非线性模型辨识方法

论文提出了单相光伏并网逆变器NARX模型的系统辨识方法。针对商用光伏并网逆变器的“黑箱”特征，以及现有的线性化建模方法无法解决逆变器的强非线性问题，将单相光伏并网逆变器视为“黑箱”，无需逆变器内部电路、功率开关器件等拓扑结构和参数及其控制系统的类型和逻辑关系，仅仅利用逆变器输入–输出两侧的外部测量数据，基于NARX模型非线性系统辨识技术，可建立较为准确的数学模型，实现对商用光伏并网逆变器准确描述其动态特性的可能。辨识所得单相光伏并网逆变器NARX模型结构简单，运算量小，在模型的复杂性和模型的精确性方面取了很好的平衡，适用于电力系统对并网光伏发电系统的调度、联合运行与协调控制、随机模拟等需要快速建模与简单模型结构的研究领域。 The system identification method of single-phase photovoltaic grid-connected inverter NARX model was proposed. For the black box feature of commercial photovoltaic grid-tied inverters, as well as the strongly nonlinear problem of the inverter which cannot be solved by existing linear modeling approach, in this method, the inverter was considered as a black box, wherein it was not necessary to know the topology and the parameters of the inverter internal circuits and power switching devices, as well as the type and logical relations of the control system. It only used the input-output external measurement data of the inverter, based on the NARX model nonlinear system identification techniques, to create an accurate mathematical model. The model can accurately imitate the behavior of the commercial inverter, and has simple structure and a small amount of computation. It takes a good balance between the complexity of the model and the model accuracy. It is suitable for power system with the grid-connected photovoltaic system scheduling, joint operation and coordinated control, and stochastic simulation research areas, in which the fast modeling and simple model structure are required.

Qualitative: Open Source Python Tool for Quality Estimation over Multiple Machine Translation Outputs

Abstract “Qualitative” is a python toolkit for ranking and selection of sentence-level output by different MT systems using Quality Estimation. The toolkit implements a basic pipeline for annotating the given sentences with black-box features. Consequently, it applies a machine learning mechanism in order to rank data based on models pre-trained on human preferences. The preprocessing pipeline includes support for language models, PCFG parsing, language checking tools and various other pre-processors and feature generators. The code follows the principles of object-oriented programming to allow modularity and extensibility. The tool can operate by processing both batch-files and single sentences. An XML-RPC interface is provided for hooking up with web-services and a graphical animated web-based interface demonstrates its potential on-line use.

Sentence-level ranking with quality estimation

Starting from human annotations, we provide a strategy based on machine learning that performs preference ranking on alternative machine translations of the same source, at sentence level. Rankings are decomposed into pairwise comparisons so that they can be learned by binary classifiers, using black-box features derived from linguistic analysis. In order to recompose from the pairwise decisions of the classifier, they are weighed with their classification probabilities, increasing the correlation coefficient by 80 %. We also demonstrate several configurations of successful automatic ranking models. The best configurations achieve a correlation with human judgments measured by Kendall's tau at 0.27. Although the method does not use reference translations, this correlation is comparable to the one achieved by state-of-the-art reference-aware automatic evaluation metrics such as smoothed BLEU, METEOR and Levenshtein distance.

Activated sludge characteristics affecting sludge filterability in municipal and industrial MBRs: Unraveling correlations using multi-component regression analysis

In the field of membrane bioreactors (MBRs) many membrane fouling related questions still remain unanswered. The goal of this research is to unveil some of the black-box features of activated sludge filterability by correlating the results from activated sludge filterability measurements following the Delft Filtration Characterization method (DFCm) with a large set of activated sludge characteristics. Ten different MBRs in Belgium and the Netherlands were sampled in both winter and summer. All samples were subjected to the DFCm, automated image analysis and an extensive set of standardized measurements. No clear correlation could be found between a single sludge parameter and activated sludge filterability. However, a combination of sludge morphology and relative hydrophobicity (RH) allows for a clear classification of activated sludge into two classes, i.e., bad and poor to good, implying that deflocculation and a low RH have a negative impact on activated sludge filterability. Furthermore, for sludge samples having poor to good filterability, accurate estimations of sludge filterability can be made when including more parameters. The main conclusion is that filterability can be predicted by analyzing the bioflocculation state of the activated sludge.

PAT: A pattern classification approach to automatic reference oracles for the testing of mesh simplification programs

Graphics applications often need to manipulate numerous graphical objects stored as polygonal models. Mesh simplification is an approach to vary the levels of visual details as appropriate, thereby improving on the overall performance of the applications. Different mesh simplification algorithms may cater for different needs, producing diversified types of simplified polygonal model as a result. Testing mesh simplification implementations is essential to assure the quality of the graphics applications. However, it is very difficult to determine the oracles (or expected outcomes) of mesh simplification for the verification of test results. A reference model is an implementation closely related to the program under test. Is it possible to use such reference models as pseudo-oracles for testing mesh simplification programs? If so, how effective are they? This paper presents a fault-based pattern classification methodology called PAT, to address the questions. In PAT, we train the C4.5 classifier using black-box features of samples from a reference model and its fault-based versions, in order to test samples from the subject program. We evaluate PAT using four implementations of mesh simplification algorithms as reference models applied to 44 open-source three-dimensional polygonal models. Empirical results reveal that the use of a reference model as a pseudo-oracle is effective for testing the implementations of resembling mesh simplification algorithms. However, the results also show a tradeoff: When compared with a simple reference model, the use of a resembling but sophisticated reference model is more effective and accurate but less robust.

Teaching project management using Excel

In this paper, we present an application of Excel spreadsheets in teaching the traditional Program Evaluation and Review Technique/Critical Path Method. This spreadsheet approach can be used in class with little cost in time or effort because it eliminates the need for tedious calculations and extensive algebraic manipulations, yet the students can understand and apply the algorithm without the black-box features that characterise most of the commercial project management software. To maintain a direct relationship with the traditional method, the spreadsheet uses the same familiar forward and backward operations required in the traditional manual calculations to find the critical path. Similar spreadsheet models are combined with the Solver in Excel to perform cost analysis by linear programming.

Online Fault Detection in Virginiamycin Production

It is difficult to measure online substrate, biomass, and product concentrations, due to the lack of reliable sensors in the fermentation. In view of this, the pH, dissolved oxygen (DO) concentration, and CO2 production, among others, are usually utilized in bioprocess analysis. With these easily obtained online measurements, it is possible to reconstruct the evolution of the state variables or estimate the bioprocess parameters. Neural networks, which rely on the efficacious nonlinear multivariate analysis capacity and its favorable black-box feature, are most widely applied to bioprocess analysis and fault detection. In this study, an artificial autoassociative neural network (AANN) has been used online to detect deviations from normal antibiotic production fermentation with ordinary state variables. To improve the efficiency of extracting hidden information contained in multidimensional state variables, and finally to render the AANN adequate for fault detection, we have explored the following methods: preprocessing of the data that involved normalizing the training data of the AANN; evaluation of the data that involved assessing the output of the AANN; and selection of state variables. A method for fault detection for virginiamycin production by Streptomyces virginiae was developed.