Search

Machine learning, and more recently, deep learning, have become a standard option for profiling side-channel analysis (SCA) to evaluate the worst-case security. Machine learning-based SCA has advantages over previous approaches like the template attack [1], especially in practical settings where the number of training traces is limited. The advantages of deep learning-based approaches are even more pronounced as such techniques can break protected implementations without feature selection and by using relatively small models (neural networks), [2]. However, the use of popular device leakage models brings in the issue of imbalanced datasets. For instance, Hamming weight or distance model follows a binomial distribution resulting in significantly more training samples in central classes. Further, evaluating the performance of machine learning- based SCA with standard machine learning metrics like accuracy can be misleading. Unfortunately, this problem is not trivial to circumvent by “just” using the SCA metrics as the training process with them is difficult.

Traditionally, electrocardiogram (ECG) signals are recorded and monitored over a period of time and finally analyzed by an expert. Automatic classification of cardiac arrhythmias has the potential to improve diagnostics. In this work, we explore the use of representation learning from ECG signals for cardiac arrhythmia classification. The dataset consisting of five cardiac rhythm types was created from the CPSC, CPSC-Extra, and The Georgia 12-lead ECG Challenge databases. We use a sophisticated deep learning approach for representation learning and classification, namely a combination of a Convolutional Auto-Encoder (CAE) and a Long Short-Term Memory (LSTM) classifier. CAE was used to compress the input signal that serves as input to the LSTM classifier. We also implemented a CAE-based data augmentation approach to balance the data distribution. The classification results reaching above 90% accuracy show that the use of the complex deep learning approach is suitable for addressing the problem.

This March 2022 issue (Vol. 30, No. 1) of CIT. Journal of Computing and Information Technology brings four papers from the areas of machine learning, risk prediction, anomaly detection, and social networks.

As implied by many and explicitly stated by Florida (2002), „human creativity is the ultimate economic resource.“ However, just like human knowledge, which is one of the preconditions for creativity to occur, creativity does not happen out of thin air. It has its foundations, triggers and constituents. Consequently, it can be developed and fostered as well as studied and analyzed. The aim of this paper is to propose the idea of education-based situated creativity. According to Nonaka and Zhu (2012), situated creativity assumes “sense-making, context-interpreting, situation-framing, meaning-giving, fact-projecting, problem-shaping, opportunity- exploring, future-envisioning” (pp. 172-173). Expanding the concept by adding “education” we focus on the necessity of perceiving creativity as a backbone of any educational process and its systematic honing by focusing on one of its fundamental academic preconditions – interdisciplinarity. This particular precondition is related to the cognitive mechanism of conceptual integration responsible for human creative spark (Fauconnier and Turner 2002, Turner 2014). In other words, evolutionarily speaking, human creativity is quite likely to have developed at the time when our working memory could simultaneously hold and juggle two elements that are different, distant or seemingly unrelatable (e.g. think of William Kamkwamba’s wind turbine made out of bike parts and materials from a scrapyard, or the invention of the axle that combines a rod and wheels, or the system of Braille alphabet consisting of three-dimensional bumps). We shall exemplify the idea of education- based situated creativity by discussing the content as well as the process of creating a new standard of occupation and qualification (Applied Cognitive Science/Scientist) that is oriented towards creative economy, especially various aspects of R&D, while being anchored in the hexagon of interdisciplinary study of human mind known as cognitive science.

Many solutions display DICOM images locally or in a web browser but require that the full DICOM images be transported over the network or on a physical medium. As DICOM image size and the number of DICOM images increase, their transport over the Internet becomes more difficult. Contemporary computer resources used by physicians cannot keep up with the ever-increasing amount of new DICOM data, which complicates and prolongs diagnosis. This problem is exacerbated in the COVID crisis, making it difficult to examine patients, communicate between patients and physicians, and have physical access to DICOM images and volumetric data. In this paper, we present a distributed architecture for remote rendering of DICOM images. We propose an innovative web-based solution hierarchically divided into several levels that can be used to manage patient data, DICOM images, and radiology reports based on those images. The architecture is organized to leverage existing hardware, storage, and telecommunications infrastructure. We have approached the solution in such a way that the images are stored in healthcare facilities, preferably where they were taken, and only the necessary data is securely and seamlessly transported to the physicians. Our carefully engineered solution shows significant improvements in terms of rendering time and speed of data transfer, which is about 110 ms for an image with a resolution of 1996 by 2457 pixels, achieving seamless user experience for at least Full HD images.

The deep learning approach for improved disorder classification is increasingly used in clinical decision support systems as databases have become more accessible and representative. In this study, we implemented a deep learning model for classifying atrial fibrillation (AF) and normal sinus rhythm (NSR) from multichannel ECG records. Following previous work, we chose the feedforward neural network model consisting of seven dense layers, which has been shown to be effective in classifying AF. A publicly available 12-lead ECG database, the Physionet/Cinc 2020 Challenge, was chosen because it contains clinical data from multiple sources. A balanced dataset of 2603 patients with AF and 3000 subjects with NSR were selected. Additionally, 240 synthetic AF ECG records were used for comparison. The records were preprocessed by filtering, removing DC component, normalizing, and extracting R peaks using the Pan-Tompkins algorithm. RR segments were extracted and resampled to equal size and used as inputs for the deep learning model. The dataset was split patient-wise into training (60%), validation (20%) and test (20%) sets. To take advantage of the multichannel ECG, the voting for majority class was applied to the model. The best F1-score was 95.8% (aVR lead) using the clinical database. For the synthetic database, the best F1-score was 99.8% (II lead). By applying voting to all leads, test results on clinical database were improved by 1.1%. Training the model on the clinical database showed good results (F1-score = 98.3%) when tested on the synthetic database. However, training on the synthetic database gave poor results (F1-score = 64.3%) when tested on clinical database. This study showed that we can achieve better results using multichannel classification. We also showed limitations of using the synthetic database in generalizing ECG changes in morphology during AF episodes.

Gender differences in traffic are generally analyzed through the number of accidents reported to the police. Our research aims to observe gender differences in electroencephalographical signals (EEG) while participants are driving. Time-frequency domain features and recurrence quantification analysis (RQA) features are calculated in order to analyze the differences. To compare male and female drivers, we used Mann–Whitney U test and compared correlations between features and brain regions. Female drivers showed significantly higher beta relative power in the occipital right region and significantly higher alpha relative power in the frontal regions, while male drivers showed significantly higher theta relative power in all regions except in the front right region. Most RQA features show a significant difference between male and female drivers. Also, male drivers showed significantly higher correlations between the RQA features, especially between different brain regions. These results could reflect the differences in the information processing strategies of male and female drivers, e.g. they tend to focus on different information when performing the task. That could account for reported gender differences in the number of traffic accidents and traffic behavior.

Pregled područja strojnog učenja, definicija i primjene. Pristupi strojnom učenju. Proces dubinske analize podataka. Najpoznatije tehnike strojnog učenja. Alati za provedbu strojnog učenja, dubokog učenja i analize velikih skupova podataka.

Sustavi za potporu odlučivanju u medicini. Pristupi računalnog predstavljanja znanja. Primjene dubokih neuronskih mreža u sustavima umjetne inteligencije u medicini.

Web ontology language (OWL), used in combination with the Protégé visual interface, is a modern standard for development and maintenance of ontologies and a powerful tool for knowledge presentation. In this work, we describe a novel possibility to use OWL also for the conceptualization of knowledge presented by a set of rules. In this approach, rules are represented as a hierarchy of actionable classes with necessary and sufficient conditions defined by the description logic formalism. The advantages are that: the set of the rules is not an unordered set anymore, the concepts defined in descriptive ontologies can be used directly in the bodies of rules, and Protégé presents an intuitive tool for editing the set of rules. Standard ontology reasoning processes are not applicable in this framework, but experiments conducted on the rule sets have demonstrated that the reasoning problems can be successfully solved.

Analiza biomedicinskih vremenskih nizova (BVN) obuhvaća široko interdisciplinarno područje. Istraživači u tom području nastoje kvantizirati složenost biološkog sustava kako bi se ostvarilo lakše razlikovanje između zdravog i bolesnog organizma. Uloga računarstva u tim istraživanjima je izlučivanje i dubinska analiza značajki. Cilj ove disertacije je sistematizirati pristupe izlučivanju značajki i predložiti sveobuhvatan postupak analize podataka na putu od baze podataka do konačnog modela koji ce omogućiti što bolje rezultate pri razvrstavanju poremećaja BVN. U tu svrhu predlaže se implementacija računalnog radnog okvira koji omogućuje izlučivanje velikog broja značajki za pojedinu vrstu BVN. U okviru disertacije implementacija okvira provedena je za područje analize srčanog ritma. Pritom su detaljno opisane vremenske, frekvencijske, vremensko-frekvencijske i nelinearne značajke korištene u tom području. Razvijena su dva nova nelinearna postupka za izlučivanje značajki iz BVN: abecedna entropija i napredna analiza slijednog trenda, za koje se pokazuje da poboljšavaju rezultate razvrstavanja poremećaja. Sustavni postupak vrednovan je na dva zasebna problema razvrstavanja, a postignuta je viša točnost od sličnih pristupa navedenih u literaturi. Rezultati dobiveni ovom disertacijom unaprijeđuju područje računalne analize BVN jer daju radni okvir temeljen na najboljoj praksi za postizanje visoke točnosti razvrstavanja poremećaja. Analysis of biomedical time-series encompasses a broad interdisciplinary area of scientific research. Researchers in this area intend to quantify the complexity of biological system in order to facilitate the differentiation between healthy and unhealthy organisms. The role of computing in this research is mostly related to feature extraction and data mining. The aim of this thesis is to provide a systematic approach starting from databases and continuing through feature extraction, data mining, and model evaluation that would ensure more accurate classification results for biomedical time-series analysis. As a part of the systematic approach, an implementation of a computer framework is proposed that consists of a large number of domain-relevant features. In this thesis, the framework is built for the domain of heart rhythm. The described features include: time domain, frequency domain, time-frequency, and nonlinear measures. Additional two novel methods for feature extraction are also proposed: alphabet entropy and advanced sequential trend analysis, which are shown to increase classification accuracy. The proposed method is evaluated on two distinct classification problems and the obtained results are better than any previously reported in literature. The results of this thesis contribute to the field of computer analysis of biomedical time-series because they provide a framework based on best practice for achieving high discrimination of different types of disorders.

The aim of this work is the identification of software project quality performance measures that would enable valid comparison and ranking of the completed projects. Software projects can be characterized by a set of influence factors. A subset of the influence factors is relevant for software project quality. In order to predict and determine the ranking of software projects by their success, and thus present a valid software project quality performance measure, we employ Grey system theory. Grey relational analysis is a kind of method which enables determination of the relational degree of every factor in the system. The method can be used for systems that are incompletely described with relatively few data available, and for which standard statistical assumptions are not satisfied. Relational degree between seven relevant software project quality influence factors is calculated for a set of ten software projects. The results demonstrate the usefulness and applicability of the Grey system theory in software project quality assessment.

Die Verantwortung für Fehler in einer automatisierten Fertigung liegt in der Regel beim Überwachungspersonal. Eine Produktion, in der Solldaten vollständig mit den Ist-Ergebnis sen übereinstimmen, ist nicht zu erlangen. Trotz der inhärent kleinen Unterschiede von Zielspezifikationen an automatisierten Maschinen bestimmen intrinsisch komplizierte und verkettete Einflüsse das Fertigungsergebnis. Fehlerschätzungen basieren deswegen nur auf Prozessinformationen. Eine vereinfachte eindimensionale (1D) signalbasierte Carnap’sche Auskunftsmessung der Prozesssignale wird vorgeschlagen. Die Vergleichsanalyse der Shannon’schen und Carnap’schen Auskunftsmessungen wird dargestellt und auf einer zwei stufigen automatisierten Laboranlage erprobt. Es wird festgestellt, dass von 30 Messpunkten die Shannon’sche Auskunftsmessung in 15% der Fälle fehlerhaft, dass die Carnap’sche 2D-Auskunftsmessung in 10% der Fälle zu zeitaufwendig, und dass die Carnap’sche 1D-Aus kunftsmessung annehmbar für dynamische Beschreibung von Produktionsfehlern ist.

Biomedical time-series (BTS) such as cardiac rhythm, electrocardiogram, electroencephalogram, etc., usually require in- depth analysis in order to determine the presence of disorder. Normal pattern of a particular BTS often possesses highly complex behavior and contains nonstationarities as a result of background biological processes. Modelling of normal patterns as well as disorders is troublesome because of the indefinite feature space – as any characteristic of the time-series might be considered a feature. A usual approach to determining which feature of the time-series should be analyzed is through an informed decision by a medical professional. This decision is somewhat arbitrary because in some cases there are no clear guidelines to which feature should be considered best for modelling of a particular BTS pattern. Nonlinear features of BTS have been recently developed such as: approximate entropy, sample entropy, spectral entropy, correlation dimension, spatial filling index, fractal dimension, and many others, which aim to better describe both the normal pattern as well as to distinguish normal patterns from disorders. To our knowledge, there is no freely available tool for extraction of many nonlinear features from BTS. This work aims to promote HRVFrame, a Java based open-source framework that allows users to extract a large number of nonlinear features (and a number of standard linear features) from BTS. Currently, HRVFrame is limited to feature extraction from cardiac rhthym, but upgrades for other BTS are planned. HRVFrame enables supervised learning and facilitates accurate model construction by extracting feature vectors to files that can be analyzed by standard data mining tools.

Ontology represents explicit specification of knowledge in a specific domain of interest in the form of concepts and relations among them. This paper presents a medical ontology describing the domain of heart failure (HF). Construction of ontology for a domain like HF is recognized as an important step in systematization of existing medical knowledge. The main virtue of ontology is that the represented knowledge is both computer and human-readable. The current development of the HF ontology is one of the main results of the EU Heartfaid project. The ontology has been implemented using Ontology Web Language and Protégé editing tool. It consists of roughly 200 classes, 100 relations and 2000 instances. The ontology is a precise, voluminous, portable, and upgradable representation of the HF domain. It is also a useful framework for building knowledge based systems in the HF domain, as well as for unambiguous communication between professionals. In the process of developing the HF ontology there have been significant technical and medical dilemmas. The current result should not be treated as the ultimate solution but as a starting point that will stimulate further research and development activities that can be very relevant for both intelligent computer systems and precise communication of medical knowledge.

The aim of this study was to investigate the usefulness of the undergraduate grade point average in prediction of scientific production of research trainees during their fellowship and later in career. The study was performed in 1,320 research trainees whose fellowships from the Croatian Ministry of Science, Education and Sports were terminated between 1999 and 2005. The data were analyzed using logistic regression. The results indicated that undergraduate grade point average was negatively associated with scientific productivity both during and after the fellowship termination. Other indicators, such as undergraduate scientific productivity exhibited much stronger positive association with scientific productivity later in career and should be given more weight in candidate selection process in science and research.

The paper demonstrates a nonlinear signal processing method based on an approach found in intelligent data mining. ECG signals were used as an interesting and readily available representative nonlinear domain. These signals were fed in an innovative software platform for feature extraction based on chaos theory. The resultant files were loaded into an open source machine learning software for clustering and classification analysis. The results depict 78% clustering and around 90% classification accuracy rate, which is quite impressive considering the number of features involved in the study.

Uvod u opisne logike. Odnos opisnih logika i računalnih ontologija. Ontološki jezici za semantički web. Algoritmi rasuđivanja i alati za rasuđivanje nad ontologijama.

Role of quality control in industry, software engineering and government. Control schemes: measurement and actuator signals. Process time constants: material balance control, production control, quality control. The concept of quality assurance and quality culture. Control system as a simple teleonomic, purposeful, system. Kolmogorov definition of teleonomic entropy. Realization restrictions in Kolmogorov entropy calculation. Shannon entropy as system uncertainty measure. Carnap entropy measure by calculating Voronoi diagram of respective signal. Introducing one-dimensional Carnap entropy measure of quantitative signals. Pragmatic question in calculating Shannon entropy: lack of data in some data classes. Problems of Carnap 2D entropy: multiple data points, split-down of signal space, data resolution and calculation duration, interpretation of obtained results. Pragmatic questions in calculating ne-dimensional Carnap entropy of control signals: equivalency of entropy measures for slightly different signal shapes, cases of short decision intervals. Delta modulated 1D Carnap entropy. Calculation of measurement and actuator signal entropies of a two-stage laboratory heat exchanger. Entropy contents of quality control signals in highly automated ceramic tile plant.

Development of heart failure domain ontology was an important step in the realization of the HEARTFAID platform. The purpose of the ontology was to systematize and summarize in a hierarchy the vast medical knowledge related to the medical disorder of congestive heart failure. The ontology was implemented using OWL-DL language and Protégé environment. Terminological knowledge contained in the ontology forms the basis for the implementation of the explicit reasoning procedures designed to help doctors in examining, diagnosing and treating patients. A single, large ontology had been developed for this purpose. Knowledge had been gathered from various sources: medical experts, congestive heart failure guidelines, project’s medical reports, Internet medical sites (UMLS, Mayo clinic, OpenClinical). The ontology contains a total of 200 different classes, 100 properties and relations between classes and about 2000 individual concepts. It is a precise, voluminous, portable, modular and upgradable representation of the heart failure domain.

Prediction of time series data for chaotic and web driven business transactions. Prediction technologies: linear regression, artificial neural networks, genetic algorithms and Quacol algebra. Dilemma in prediction technique: functional or stochastic model. Fit composition of prediction functions – Quacol predictor model. Elaboration on rank exclusivity, on continuity of n-point graph and theorem on sign independent algebraic operations. Practical prediction data from chaotic behavior in a ferroresonant circuit. Elaboration on prediction of web driven trading process. Modeling prediction error in Quacol algebra using triangle inequality.

Prediction of time series data. Comparison between chaotic and web driven business transactions. Technology of artificial neural networks, genetic algorithms and Quacol algebra. Takens equation for sampling, width and prediction horizon. Prediction technique: functional or stochastic model. Fit composition of prediction functions. Predicting data from chaotic behavior in a ferroresonant circuit. Prediction of web driven trading process. Prediction error: sampling interval, correlation, time horizon. Modeling prediction error in Quacol algebra using triangle inequality. Discussion of practical results.

Ontologies of the dynamic distributed Web enterprises differ. Ontology engineers need a tool for ontology updating that is swift and simple. Therefore a simplistic distributed semantic web model (SDSW) is proposed. Added to the model is an extension of the primary logic operators. Logic operators are viable on domain variables and describe the formation, evolution and updates of the ontologies with conflicting information. Knowledge Base over a specific conflict is formed that uses expert information in order to cover for the logic consistency. The measure of logic consistency is obtained which is based on the extension set of the primary logic operators. Algorithms for checking consistency are thus drastically shortened in the case of large number of object ontology variables.

U radu je opisan rad srca i princip rada elektrokardiografa. Prikazane su značajke elektrokardiograma. Navedeni su izvori iz kojih se dobivaju elektrokardiografski signali. Prikazan je dio teorije kaosa koji je primjenljiv na EKG signal. Proučavane su značajke kaosa: fazni prostor, atraktori, Ljapunovljevi eksponenti, indeks prostorne popunjenosti, korelacijska dimenzija, mjera središnje težnje, približna entropija, test na determinizam. Za potrebe izlučivanja značajki izrađen je program "ECG Chaos Extractor". Program uspješno izlučuje značajke kaosa iz EKG zapisa te ih predaje Weka sustavu za klasifikaciju. Na temelju malog broja primjera ne može se dati konačna ocjena uspješne primjenljivosti teorije kaosa na EKG signale.