Information and Software Technologies

Due to recent advancements in technology, developers of intrusive software are finding more and more sophisticated ways to hide the existence of malicious code in software environments. It becomes difficult to identify viruses in the infected data sent in this way during analysis and detection phase of malware. For this reason, a significant amount of consideration has been devoted to research and development of methodologies and techniques that can identify miscellaneous malware without compromising the execution environment. In order to propose new methods, researchers are investigating not only the structure of malware detection algorithms, but also the properties that can be extracted from files. Extracted features allow malware to be detected even when virus creation tools change.

The authors of this study proposed a data structure consisting of 486 attributes that describe the most important file characteristics. The proposed structure was used to train neural networks to detect viruses. A set of over 400,000 infected and benign files were used to build the data set. Various machine learning algorithms based on unsupervised (k-means, self-organizing maps) and supervised (VGG-16, convolutional neural networks, ResNet) learning were tested. The performed tests were designed to determine the usefulness of the tested algorithms to detect malicious software.

Based on the implemented experimental research, the authors created and proposed a neural network architecture consisting of Dense and Dropout layers with L2 regularization that enables the detection of 8 types of malware with 98% accuracy. The great advantage of the article is the research carried out based on a large number of files. The proposed neural network architecture recognizes malware with at least the same accuracy as solutions offered by other authors and can be practically used to protect workstations against malicious files.

Widespread use of cloud computing resources calls for reliable network connections, while anomalies in network traffic impact the availability of cloud resources in a negative way. Anomaly detection tools are essential for identifying and forecasting these network anomalies. In recent years machine learning methods are gaining popularity in implementations of anomaly detection tools. Given the variety of network anomaly types and the availability of diverse machine learning algorithms, developers of anomaly detection software and administrators of cloud infrastructures are presented with a wide range of possible solutions.

This article presents a survey of the most popular machine learning methods that are applicable to detecting anomalies in cloud networks. In order to be able to classify and compare these methods, six major criteria (training approach, training time, preferred areas of application, discovery of unprecedented anomalies, dataset’s influence on anomaly prediction and problem of vanishing or exploding gradient) are discerned and discussed in detail, providing their implications on the evaluated methods. Subsequently, the criteria are used to review the features of the main machine learning methods for anomaly detection and to provide insights about using the methods to identify abnormal network behavior.

The last part of the study lists the examined machine learning methods and appropriate tools for anomaly monitoring and detection. The provided lists are then used to draw final conclusions that provide the recommendations for employing the aforementioned algorithms and tools in various cases of anomaly detection.

System monitoring is crucial to ensure that the system is working correctly. Usually, it encompasses solutions from the simple configuration of static thresholds for hardware/software key performance indicators to employing anomaly detection algorithms on a stream of numerical data. System logs, on the other hand, is another golden source of the system state, but often it is overlooked. Combining system logs with load metrics could potentially increase the accuracy of anomaly detection. We propose a robust pipeline and evaluate several of its variants for solving such a task at scale and in real-time. Experiments with proprietary logs from an enterprise Kafka cluster reveal that pre-processing with an autoencoder prior to applying the isolation forest method can significantly improve the detection performance.

Extraction of new knowledge from earlier obtained and integrated knowledge is one of the main stages of intelligent knowledge analysis. To handle such a task, a knowledge-based system should be able to decompose complex or composite knowledge structures and extract new knowledge items, which were hidden or non-obvious before. Existed approaches to decomposition within object-oriented paradigm provide different variants of partitioning or fragmentation of main knowledge structures, such as objects, classes, and relations among them, however, most of them do not consider semantic structural and functional dependencies among properties and methods of classes that affect on the decomposition process. In this paper, we introduced concepts of fuzzy structural and functional atoms, as well as molecules of fuzzy homogeneous classes of objects, within such a knowledge representation model as fuzzy object-oriented dynamic networks. In addition, we proposed the algorithm for the decomposition of fuzzy homogeneous classes of objects, which implements the idea of universal decomposition exploiter of fuzzy classes of objects, and constructs semantically correct subclasses of a fuzzy homogeneous class of objects by solving appropriate constraint satisfaction problem that defines decomposition conditions. To demonstrate some possible application scenarios, we provided an appropriate example of the decomposition of a fuzzy homogeneous class of objects.

Audio processing technology is happening everywhere in our life. We ask our car to make a call for us while driving, or we let Alexa turn off the light for us when we don’t want to get out of bed before sleep. In all of these audio-based applications and research, it is AI and ML that makes the computer or the smart phone understand us via our voice [1]. As an important part of artificial intelligence (AI), especially machine learning (ML), which has had great influences in many areas of AI and ML-based research and applications. This paper focuses on deep learning structures and applications for audio classification. We conduct a detailed review of literature in audio-based DL and DRL approaches and applications. We also discuss the limitation and possible future works for audio-based DL approach.

This research aims to provide an overview of the technological solutions of instant payment with cryptocurrencies in the tourism sector. The analysis has been done on various cryptocurrencies and technologies related to them, their transaction processing speed, alternatives on how to optimize it, and what impact it has on cryptocurrency payments adoption to the tourism sector. It was noticed that various places of services in the tourism sector could take a lot of benefits from accepting the cryptocurrency transactions, such as an additional group of clients, eased payment mobility and fast settlement of the received funds, but they are also avoiding accepting the most popular cryptocurrencies due to the volatility of their market, high transactional costs and low transaction confirmation speed, which would cause an issue to manage the customers’ queue while comparing to traditional currency payments. This research revealed that there are solutions to avoid the mentioned drawbacks by offering less popular cryptocurrency acceptance or taking the possible risks of accepting 0-confirmation transactions. The prototype of instant payments solution is suggested where cryptocurrency payment adoption in the tourism sector could be reached by accepting 0-confirmation transactions of the most popular cryptocurrencies with an instant exchange to fiat currencies at the merchant side.

Human body segmentation is an intermediate step in many applications. Current state of the art shows the best results of segmentation when deep neural networks are used, however, they require lots of annotated data to learn from. Annotating data for segmentation is a very tedious process since part of the image has to be marked as foreground. This involves much higher amount of manual work by a human than classification where only a correct label must be picked. Geometrical solutions may be used to assist the human, but their mistakes must be corrected manually. The goal of this research is to reduce the total time of segmentation annotation process. This is achieved by a machine learning solution that improves the accuracy of an existing geometric algorithm when applied to human body segmentation. It is trained from 8 introduced metrics acquired after point cloud split based on 2-Means cut. The approach has been trained on two real life datasets that include humans in different positions. Observed results show that accuracy is improved in the most complex scenes at a performance penalty. However, this is a good trade-off in case the original algorithm is unusable due to very low accuracy.

The amount of data is growing at an extraordinary rate each year. Nowadays, data is used in various fields. One of these areas is economics, which is significantly linked to data analysis. Policymakers, financial institutions, investors, businesses, and households make economic decisions in real-time. These decisions need to be taken even faster in various economic shocks, such as the financial crisis, COVID-19, or war. For this reason, it is important to have data in as frequent a range as possible, as only such data will reliably assess the economic situation. Therefore, automated systems are required to collect, transform, analyse, visualise, perform other operations, and interpret the results. This paper presents the concept of economic activity, classical and alternative indicators describing the economic activity, and describes the automated economic activity monitoring system. Due to the different economic structures and the different availability of data in different countries, these systems are not universal and can only be adapted to specific countries. The developed automated system uses working intelligence methods to predict the future values of indicators, perform clustering, classification of observations, or other tasks. The application’s developed user interface allows users to use different data sources, analyses, visualisations, or results of machine learning methods without any programming knowledge.

BIM6D is an aspect of building information modeling (BIM) that allows for a detailed analysis of a building's energy performance in order to improve energy and light efficiency, which in turn leads to a more sustainable building utilization. Predictions of a building's energy consumption can have added value in different aspects and for different building actors, be they engineers, architects or the building users themselves. The objective for this study is to explore mathematical and artificial intelligent approaches for predicting thermal energy consumption in buildings and to examine its use for BIM6D. The dataset used in the research includes several years of hourly thermal energy consumption collected in one block of Kaunas city. Experiments have been carried out using different forecasting methods. In terms of prediction accuracy, it is worth highlighting the Extra Trees with \({MAE < {3}{\text{.5}}\;{\text{kWh}}}\) and Support vector regression (SVR) with \({MAE \le {2}{\text{.63}}\;{\text{kWh}}}\). However, Extra Trees seems to be the best in terms of MAPE (38.65%). Although prediction time is not the most critical parameter, it should be noted, that Extra Trees, SVR and auto-regressive models were found to be the most time-consuming (from 2 to 4 min) to linear models (<1 s) and extreme gradient boosting (~3 s) and that these results may influence the selection of a model for real-life operation.

The Web is characterized by user interaction on Online Social Networks, the exchange of content on a large scale, and the presentation of one’s own life on several digital channels using different media. Users strive to reach as many people as possible with their content while also distributing traffic across the various networks. To simplify this, there are Social Media Reference Landing Pages where users can bring together their numerous social media profiles. Our research project investigates the threat to users posed by the shared content, such as blackmailing or doxing. An important step is finding and merging different user profiles, primarily based on hints, similar user names, or links. In this paper, we show how Reference Landing Pages make it easier to create comprehensive Digital Twins, which we can use to compute and make tangible the risk of thoughtless sharing of information to users.

The Internet of Things (IoT) has combined the hardware components with software elements by providing users with remote control and management facilities. From safety-critical systems to security devices and industrial appliances, every appliance makes use of IoTs. Whereas security issues such as SQL injections, Denial of Service/Distributed Denial of Service (DOS/DDOS) attacks, the forged transmission of messages, or man in the middle (MITM) are major security threats among smart devices. Any purging of data causes privacy issues while the subsequent assessments made using modified information are also erroneous. This security hole needs comprehensive non-cryptographic data-security techniques and frameworks which would help developers in creating secure systems on heterogeneous devices. Algorithms like blowfish and Data Encryption Standard (DES) do not have the uniquity which AES does, making them more vulnerable to attack this research paper focuses on the communication security issues in IoT systems. We have proposed an End-to-End Encryption using AES in IoT (EAES-IoT). Validation of the proposed algorithm has been done in a case study of the Smart Voice Pathology Monitoring System (SVPMS) by sending the encoded data to the application layer through Application Programming Interface (API). We compared results to ensure the authenticity of the data and they were found promising. Data access is provided only to authorized individuals by providing a shared key for decryption of the alphanumeric string of data shared between devices. The proposed algorithm will provide future directions to meet security challenges in the IoT.

The application of virtual reality solutions for rehabilitation is hard, because stroke patients usually suffer motor, gait, and visual field impairments. This article discusses important aspects that should be addressed when developing programs of a similar nature. The system for stroke patients’ cognitive rehabilitation is introduced. The development process of the system is described in detail. The results were validated using the Content Validity Index (CVI). The validation results revealed that the tasks created are suitable for stroke patients’ cognitive rehabilitation.

Regarding IoT applications, the efficiency has immensely upgraded, though the product features remain the same, the progress in extremely low power sensing and computing has boosted the efficiency and thereby the power consumption of IoT devices have drastically dropped. With this change happening for the first time in history, it is actually feasible to tap into this appreciable energy available in our surroundings to power such electronic devices. The tapped energy from environment not only enables self-reliant electronics but also gives a chance for addition of newer features in IoT applications. This paper is devoted to the development of a multilayer PVDF based piezoelectric vibration energy harvesting device for powering wireless sensor networks and low power electronic devices. The purpose of the device is to be the power supply to endless applications of information technology.

The designed energy harvester successfully generates an average power of \(9.2\,\upmu {\text{W}}/{\text{g}}/{\text{mm}}^{3}\) with a resonant frequency of 43 Hz, generating at least 15 V rms voltage and 495 μW power for acceleration 1 g. The commercial piezo sensors generate power of only \(10\,{\text{nW}}/{\text{g}}/{\text{mm}}^{3}\). This work reveals the challenges and limitations involved in constructing a realistic piezoelectric energy harvesting system and how to overcome them with the proposed harvester design. The method of fabrication and design of the proposed energy harvester are also discussed. Comparison of the harvester results with other author works is presented. Future recommendations, suitable application areas and market size information is also provided.

One of the problems with conventional conceptual modelling methods is that they do not take into account certain important semantic interdependency types between the static and dynamic aspects. Integrity of dimensions is crucial for successful reasoning and solving problems that occur in conceptual modelling. Typically, conceptual modelling methods project various aspects of information systems using different graphical representations. Therefore, to reach semantic integration of various architectural aspects is very difficult. This paper presents semantically integrated conceptual modelling method. This method enables stability and flexibility of the diagrams that are very important for managing constant changes of organizational and technical requirements. It shows how alternative actions introduced into different scenarios. This is also important for controlling semantic integrity and for maintaining holistic representations of different aspects. Holistic modelling approach enables reasoning about system architecture across organizational and technical system boundaries. On a simple hotel reservation system scenario, it is demonstrated how different actions can be decomposed into more primitive underlying interaction loops. Integrated conceptual modelling method is important for evaluation of expressive power of conceptual modelling languages.

The design and implementation of vehicle suspension systems is complex and time-consuming process that usually leads to production delays. Although different Model Driven Engineering (MDE) technologies like EAST-ADL/AUTOSAR are frequently applied to expedite vehicle development process, a framework particularly dealing with design and analysis of vehicle suspension is hard to find in literature. This rises the need of a framework that not only supports the analysis of suspension system at higher abstraction level but also complements the existing standards like EAST-ADL. In this article, a Model driven framework for Vehicle Suspension System (MVSS) is proposed. Particularly, a meta-model containing major vehicle suspension aspects is introduced. Subsequently, a modeling editor is developed using Eclipse Sirius platform. This allows the modeling of both simple as well as complex vehicle suspension systems with simplicity. Moreover, Object Constraint Language (OCL) is utilized to perform early system analysis in modeling phase. Furthermore, the target MATLAB-Simulink models are generated from source models, using model-to-text transformations, to perform advanced system analysis. The application of proposed framework is demonstrated through real life Audi A6L Hydraulic active suspension use case. The initial results indicate that proposed framework is highly effective for the design and analysis of vehicle suspension systems. In addition to this, the analysis results could be propagated to EAST-ADL toolchains to support full vehicle development workflow.

The purpose of this paper is to present continuous results of the research in financial data analysis. Many organizations face challenges by processing a colossal quantity of financial data for evaluation of the current state of the organization, for analysis of future strategies and other purposes. One of the possible ways to analyse financial data is to use process mining techniques. This paper proceeds with analysis and usage of financial data cubes dimensions using General Ledger information of particular organizations in the Netherlands. The research project is funded by European Regional Development Fund according to the 2014–2020 Operational Programme for the European Union Funds’ Investments under measure No. 01.2.1-LVPA-T-848 “Smart FDI”. Project no.: 01.2.1-LVPA-T-848–02-0004; Period of project implementation: 2020–06-01–2022–05-31.

Authors developed an intelligent method of forming a consumer basket based on data from supermarket chains, which allows modifying the set of goods in the consumer basket and defining a living wage. The consumer basket is forming on a base of k-means clustering approach. The algorithmic structure of the proposed method is described. Experimental research is carried out using the Customer Personality Analysis dataset from the Kaggle platform. After data normalization and clustering, the clusters relative to the amount (USD) of purchased goods for 2 years were analyzed. As a result, the cluster (consumer basket) was selected which includes 27% of middle-aged customers of various ages and counts such goods as fish, meat, sweets, wine and equipment. The novelty of the paper is the automated and intelligent forming the set of goods in the consumer basket, which may promote survival during humanitarian and economic disasters, especially in times of economic crisis (war, pandemic).

The applications of Robotic Process Automation (RPA) are many and growing every year, covering banking and financial operations, insurance functions, auditing processes, logistics planning services and more, but automating invoice processing is still more challenging. However, different structure, the variety of keywords, the abundance of types of information makes the detection and retrieval of the information very complicated. Therefore, the currently solutions currently work well with predefined document structures. The aim of this study is to develop a solution based on RPA, Optical character recognition (OCR) technologies and deep learning methods for automated invoice processing without being bound to a specific document structure. The study focuses on five key fields of invoices that are most important to identify and read. The results showed that almost 83% of all invoices used in the experiments were processed correctly. In terms of the results for the detection of individual field information, the best results were found for the “date” and “total amount” fields, with 93.21% and 87.81% respectively, but the detection of the “seller” and “buyer” fields is complicated and requires extensive additional research. Experiments evaluating the document processing time of human the developed robot showed that the performance of human document processing decreases with the volume of documents processed, while the RPA time is almost constant. RPA is 1.76 times more efficient with 500 documents and 2 times faster with 1000 documents.

In this study, we analyze the trends of COVID-19 related communication in Croatian language on Twitter. First, we prepare a dataset of 147,028 tweets about COVID-19 posted during the first three waves of the pandemic, and then perform an analysis in three steps. In the first step, we train the LDA model and calculate the coherence values of the topics. We identify seven topics and report the ten most frequent words for each topic. In the second step, we analyze the proportion of tweets in each topic and report how these trends change over time. In the third step, we study spreading properties for each topic. The results show that all seven topics are evenly distributed across the three pandemic waves. The topic “vaccination” stands out with the change in percentage from 14.6% tweets in the first wave to 25.7% in the third wave. The obtained results contribute to a better understanding of pandemic communication in social media in Croatia.

Modern deep learning based speech recognition methods allow for achieving phenomenal speech recognition accuracy. But it requires enormous amounts of data to train such systems to achieve high recognition accuracy. Many less widely spoken languages simply do not possess the necessary amounts of speech corpora. The paper presents attempts to evaluate DeepSpeech-based speech recognition efficiency with the limited amounts of training data available and the ways to improve the accuracy. The experiments showed that the accuracy of DeepSpeech2 recognizer with about 100 h of speech corpora used for training is quite modest but the application of simple grammatical constraints allowed to reduce the word error rate to 23–25%.

Modern deep learning-based speech recognition methods allow for achieving phenomenal speech recognition accuracy. But this requires enormous amounts of data to train. Unfortunately, developers of recognizers for less widely spoken languages are often facing the problem of scarce resources to train recognizers. The paper presents a novel method to increase recognition accuracy by post-processing of the text outputs of two different speech recognizers. The method is using machine learning to find a more likely symbol or group of symbols from two different deep learning-based recognizers. The experiments showed that the method allows increasing recognition accuracy by 3%.

The paper presents a study on the relevance of the challenge-based learning (CBL) approach in today’s learning process, introduces the stages of CBL and their key elements, reviews the possible ways of learning incorporating technologies and provides key recommendations for a successful and effective learning process for responsible education. The main objective of this paper is to analyze how technology-enriched learning can ensure the successful implementation of CBL in the study process. The paper also presents good practice and the existing experience of the authors and provides a wide range of technological solutions for the application of CBL.

Open online courses are often used in formal education to provide added value for the students by helping gain new skills and competences or as extra material tasks in addition to the formal education course. The course described in this paper is developed and piloted fully open online. It is also integrated into formal education. This paper presents a case on integration of open databases course into formal education together with student feedback on the course quality and effectiveness of course delivery process.

This paper presents the national cases on the open online courses recognition. This result is directly related to the non-formal education, i.e. which means as a collective term for all forms of learning and education which happens in all fields outside of formal educational systems. The lack of accreditation is an issue to be addresses to policymakers. The lack of formal recognition for MOOCs could be a reason why they are taken by participants who already have university degrees, because they are taking courses to update their skills. The paper shows some national cases of the open courses recognition.

The aim of the paper is to present the processes of the game implementation and design. Nowadays technologies could play an active role in promoting social inclusion and equal participation by providing people with interactive experiences on these subjects. Paper authors are developing a game that will engage learners and motivate them to learn from simulated experience-enhancing critical reflection on social and political circumstances, build skills and stimulate interest for collective action. The paper presents a case study on gaming implementation in practice.

Massive Open Online Courses (MOOCs) can be especially useful in the Science, Technology, Engineering, Arts and Mathematics (STEAM) field to provide a large number of teachers with appropriate content for their professional development. On the other hand, we still have to think, how to use STEAM in virtual platforms for robotical learning (r-learning) and ensure appropriate skills attainment of teachers.

Best practices, how to design better MOOCs for teacher’s professional development, are still in search. In this paper, we offer MOOCs with a theoretical background describing STEAM models for teachers continues professional development (TCPD).

The aim of the paper is to present the relationship between proposed framework for teachers’ professional development in STEAM and MOOC platform.

The objectives of the study are: 1) to presents the analysis of the existing STEAM models for teachers’ professional development; 2) to presents framework for teachers’ professional development in STEAM Teacher Training & Training Curriculum model with robotical education; 3) to presents methodology for MOOC design on Framework for Teacher Professional Development in STEAM usage for r-learning implementation.