Abstract: Precision agriculture, which involves the use of information systems in agriculture, is becoming increasingly popular as it helps ensure optimal crop yields while maintaining plant health and protection. Our aim is to address the need for a comprehensive monitoring system for perennial plantations by developing a software solution that processes images and provides actionable insights into plant health and growth. The proposed system is an idea for a software solution based on artificial intelligence. In order to generate information about plants, a system using object detection and image segmentation approaches is proposed. Ultimately, our goal is to create a system that empowers farmers and agricultural professionals to make data-driven decisions about plantation management, leading to improved crop yields and sustainability.

Abstract: Ransomware is malicious software that encrypts or blocks access to data or computer systems, demanding a ransom for decryption or reestablishment of access. Ransomware attacks have become more sophisticated, using a variety of distribution techniques, including phishing emails, malicious ads, and exploiting vulnerabilities in computer systems. This type of attack can have serious consequences, including financial loss, damage to the reputation of organizations and violation of user privacy. This scientific paper gives a comprehensive overview of ransomware attacks, analyzes their characteristics and consequences, and explores the various defense mechanisms that are used to prevent, detect, and respond to these threats. Defense mechanisms against ransomware attacks include prevention, detection and response. Prevention focuses on implementing best practices such as employee training, system updates, and implementing security protocols. Detection involves the use of advanced tools to detect the presence of ransomware, while response involves isolating infected systems, notifying authorities and recovering data. The paper also shows examples of successful defense mechanisms that have been applied in practice. Advances in ransomware attack research and defense mechanisms are crucial for developing more effective protection strategies. Understanding the characteristics and tactics of attackers makes it possible to adjust defensive strategies, while identifying best practices contributes to improving the overall security of the digital environment. The aim of this paper is to provide a thorough overview of ransomware attacks and defense mechanisms, contributing to improving the security of organizations and individuals in the face of increasingly complex cyber threats.

Abstract: The aim of this study is to assess and evaluate different machine learning techniques, in synergy with deep learning, for predicting immersion in virtual reality. Python modules, combined with descriptive statistical tools and nntraintool in Matlab environment are used to accomplish this task. Analysis is conducted over the set of 401 cases with 23 variables. Linear, lasso, lassoCV, ridge, and Bayesian regression are tested for different sets of input data. These input data consist of 22 or 8 independent variables. The best score (0.4419295) is obtained for lasso CV model. In order to improve the results Artificial Neural Networks (ANNs) with architecture types 22-n-1, and 8-n-1, are tested. Number of neurons in hidden layer is marked with n (n = 5, 10, 20, 30, 40, 50), 22 and eight are numbers of neurons in input layer, while one stands for the number of neurons in the output layer and corresponds to the level of immersion. The ANNs are trained with Levenberg - Marquardt (LM) algorithm, and Scaled Conjugated Gradient (SCG) algorithm. The results are evaluated according to the criteria of the highest value of correlation coefficient (R) for train data and the lowest values of Mean Squared Errors. These results are cross-referenced with the best performances of proposed ANNs. Architecture 8-5-1, trained with SCG algorithm, is proven to be the best solution, with R-value for train data equal to 0.726, and the value of Mean Square Error (MSE) equal to 0.620. Best validation performance is 0.689, reached at epoch 18.

Abstract: This paper analyzes the public transportation network of Belgrade using techniques from the field of complex network analysis. Using publicly available data about the stops and routes, we built two network models of the Belgrade transportation network. The first model is based on the often-used L-space model and serves as a benchmark for the second, spatially augmented model. Spatial augmentation is done with the use of a novel station aggregation algorithm we proposed, named Spatial Coordinate-based Agglomerative Node Aggregation (SCANA). SCANA aims to optimize transit systems by identifying and grouping spatially proximate stations. The study leverages geographical and connectivity data from Belgrade’s public transportation system, employing SCANA to identify clusters of stations with high spatial coherence. The algorithm’s application facilitates the creation of aggregated station groups, enabling a more efficient representation of the complex network structure. To evaluate the benefits of spatial augmentation with the SCANA algorithm, we compare global network metrics, inspect the top nodes based on centrality measures, and explore the community structure of the two network models. We show that the spatially augmented network model demonstrates measurable improvements regarding network connectivity and ease of transport: a decrease in average path length of 11.5% and a decrease in network diameter of 13.2%. The augmented network exhibits a 2.6 times higher average clustering coefficient, and the community structure expressed by the spatially augmented model is more coherent and adheres better to the geographical and organizational constraints of the city.

Abstract: In this comprehensive scientific manuscript, the intricate process of developing and deploying cutting-edge alarms and warning systems tailored for the precise detection of physico-chemical pollution within river ecosystems is investigated. Through meticulous investigation, a spectrum of methodologies and technologies is thoroughly scrutinized, unraveling their intricate interplay and practical applications in the realm of environmental monitoring. The overarching goal is to provide nuanced and valuable insights, advancing the collective understanding of how pollution in river environments can be effectively detected and responded to. An in-depth analysis of early warning systems is offered, elucidating their potential to tackle the intricate challenges posed by physico-chemical pollution in river ecosystems. By exploring the diverse applications of these systems, substantive contributions are made to the discourse on sustainable management and preservation of critical freshwater ecosystems. Through this detailed examination, the current landscape of environmental monitoring is not only illuminated but also the way is paved for informed decision-making and strategic interventions to safeguard the health and sustainability of vital water resources.

Abstract: In recent years, there has been a notable shift towards the use of deep learning methodologies in forecasting with promising results. Specifically, models such as Long Short-Term Memory (LSTM) networks and Temporal Convolution Networks (CNNs) have demonstrated significant potential in this domain. However, the black-box nature of neural networks poses a substantial hurdle to their widespread adoption, primarily due to the lack of understanding and trust by the end users. Providing explanations of the model’s prediction should increase the trust in the system and make the peculiar decisions easily examined. This paper presents a survey of machine learning systems for time series prediction based on neural networks, augmented with eXplainable AI (XAI) techniques. A architecture integrating LSTM neural networks with SHAP (SHapley Additive exPlanations) values is proposed. The XAI module generates explanations for model predictions, which are subsequently presented to users alongside prediction results, allowing them to understand the system’s decisions. Furthermore, additional benefit of XAI module is the possibility to experiment with different prediction models and compare input feature effects.

Abstract: The paper presents the basic principles and frameworks of the ISO 10001 standard and their practical use in the information technology (IT) system. , designing, developing, implementing, maintaining and improving codes of conduct for user satisfaction. This paper aims to show how, in today's era of digitisation and real-time data processing, we can establish and use a code of conduct that focuses on user satisfaction. Before presenting the practical application of the basic principles of the mentioned standard, special attention is paid to users as the most important factor that dictates success or failure on the market, the importance and methods of collecting feedback from users and their proper processing and handling through the standard, to achieve technological, economic and social benefits. The standard allows for harmonisation and improvement of the technical specifications of IT services. It reduces the behaviour within the organisation to clearly defined principles and guidelines, which make IT services more flexible and in accordance with new user requirements without unnecessarily complicating the organisational structure and activities. In conclusion, it can be pointed out that standards shape organisations to ensure consistency, quality, safety, compatibility and reliability in different areas, industries and products. However, for the complete conquest of the user audience, it is necessary to have a certain level of knowledge and experience.

Abstract: Residential buildings are major energy consumers, with actual energy usage often exceeding predictions due to operational inefficiencies. This discrepancy leads to higher costs and increased reliance on carbon-intensive Power Plants. Improving energy efficiency is essential, and the Internet of Things (IoT) offers a promising solution for real-time energy management. However, data heterogeneity and interoperability issues prevent widespread IoT adoption in Building Management Systems (BMS). This paper presents an interoperability approach within a research project that applies Demand-Response (DR) strategies to optimize energy use in residential buildings. The proposed platform integrates open standards and communication protocols to ensure compatibility across different systems. Key features include real-time data collection from sensors and smart meters, external data integration (e.g., weather forecasts, energy pricing), and advanced analytics for energy optimization. A central component is the middleware, which facilitates seamless communication between field-level devices, external services, and visualization tools. The platform employs a Canonical Data Model (CDM) to standardize message formats and communication protocols, minimizing integration complexity in multi-vendor environments.

Abstract: Bibliometric analyses often rely on combining data from multiple citation databases to achieve comprehensive coverage and accurate results. In our previous study, we presented an open-source tool that enables automated merging and deduplication of records from Scopus and Web of Science (WoS). However, emerging approaches - like integrating OpenAlex, minimizing manual data handling, and enabling reproducible automation, are still underdeveloped. To address this, we enhance our toolkit with four key features: (i) integration of the OpenAlex API, (ii) execution of the entire workflow on GitHub Actions for unattended data retrieval, (iii) a fork-based configuration model that simplifies collaboration, and (iv) an aggregation module coupled with versioning via git tags. These improvements reduce researcher workload and advance the principles of openness and reproducibility in bibliometric practice.

Abstract: This paper presents an implementation of a service designed for managing data in distributed clouds, inspired by the concepts of hard and soft links in the Linux operating system. Each application has a hard link pointing to its own data. Additionally, an application can access data from another application if it has a soft link to the other application's data. Furthermore, small collaborative applications can be bound to links to enhance performance and data locality. The paper provides a detailed explanation of these mechanisms and offers an overview of the technical solution.

Abstract: This paper aims to describe the tool created to alleviate and speed up the process of updating static web pages of the courses held by the Department of Computer Science and Information Technology of the School of Electrical Engineering, University of Belgrade. The paper gives an overview of the current solutions and their shortcomings. The end result of this paper is a tool that automates a large portion of the process of updating static web pages and reduces the amount of user input.

Abstract: As new cloud models evolve, traditional monitoring solutions often struggle to meet the scalability, lightweight operation, and resilience requirements of modern systems. This paper introduces a hybrid monitoring system tailored for distributed cloud (DC) environments, which face the challenges of dynamic, geographically dispersed, and resource-constrained infrastructures. Our approach decentralizes both data collection and aggregation within regions, effectively reducing latency and avoiding the bottlenecks that would arise from centralizing these processes in the cloud. Aiming to minimize the overhead on individual nodes, the system offloads data collection and processing to metrics servers located within each region. Configurable filtering strategies ensure efficient data handling by discarding insignificant changes, while continuous aggregation at the region-level metrics servers improves query responsiveness. To enhance resilience, the system leverages cloud-managed replication and failure detection, ensuring uninterrupted service even during server or network failures. Real-time alerting is handled locally on metrics servers, reducing response times. The system supports aggregated views through client-defined aggregation rules, offering timely insights into system performance. Our solution is designed to scale efficiently as new nodes and regions are added, with minimal impact on network load and performance.

Abstract: Social engineering attacks exploit human psychology rather than technical vulnerabilities, making them particularly dangerous in today’s increasingly digital world. These types of cybercrimes often involve tricking individuals into giving away confidential information or performing actions that compromise their security. As social engineering tactics evolve, their sophistication increases, making it more challenging for individuals and organizations to recognize these threats. It is crucial to understand the different types of scams and the red flags that indicate a fraudulent email. Social engineering email scams continue to evolve, but with the right knowledge and precautions, individuals and organizations can defend against these sophisticated attacks. By recognizing the warning signs and staying vigilant, you can prevent falling victim to these manipulative schemes.

Abstract: This paper presents a comprehensive approach to predicting aggressive driving behaviors using a deep neural network (DNN) model. The model is trained on a diverse dataset containing both behavioral and environmental metrics, including vehicle specifications, road conditions, weather parameters, and driver interaction patterns. Key features, including vehicle speed, length, road lane, weather conditions, and preceding vehicle dynamics, provide a solid foundation for modeling driving behavior. Using carefully preprocessed data, the model predicts the driving style for each individual trip, classifying it into categories such as cautious, potentially aggressive, and aggressive. The core of the study lies in the implementation and application of a deep learning model, specifically a multilayer perceptron (MLP), which is designed to capture complex, non-linear relationships between input variables and driving styles. The architecture incorporates several hidden layers, with dropout and batch normalization techniques to enhance generalization and prevent overfitting. Additionally, hyperparameter tuning was integrated to further optimize the model's performance. The model's effectiveness is thoroughly evaluated using a range of performance metrics, achieving a promising accuracy rate of 87%. The proposed method holds significant potential for dynamically constructing driver profiles based on realtime driving behavior information, with applications in autonomous driving systems, intelligent vehicle safety technologies, and personalized insurance models, offering a powerful tool for proactive risk assessment and driver monitoring.

Abstract: The role of the flood early warning system is to provide information on forecasted water levels along river reaches obtained using mathematical models. Accordingly, it is necessary to ensure adequate initial conditions of mathematical models for the forecast period. Initial conditions are provided through the assimilation process, where corrections to input data, states, or model parameters are made based on the difference between observed and modeled values in the pre-forecast period. This paper presents the initial testing results of the proposed assimilation method for a hydrological-hydraulic model within the Kolubara Flood Early Warning System, based on solving an optimization problem using a genetic algorithm.

Abstract: Managing the reconfiguration of large-scale, geo-distributed cloud systems presents significant challenges, particularly when ensuring minimal downtime and scalability. Traditional approaches rely on centralized control mechanisms, which can become bottlenecks in modern distributed environments. As new cloud models, such as the distributed cloud (DC), emerge to meet the demands for lower latency and improved privacy, scalable and reliable reconfiguration strategies are essential. This work focuses on enhancing the existing reconfiguration architecture by reducing dependence on centralized components. We propose a hybrid approach for reconfiguring DC applications and system components that integrates a centralized method for smaller deployments with a decentralized, gossip-based, method, for large-scale DCs. Our solution enables efficient propagation of configuration updates across clusters while allowing selective application of changes based on node labels. Beyond reconfiguration, our approach can facilitate other cluster-wide information-sharing tasks. Our solution improves scalability, fault tolerance, and flexibility, making distributed cloud reconfiguration more efficient and practical for diverse deployment scenarios.

Abstract: Due to the increasing volume of innovations emerging daily across various domains, there is a growing need to efficiently filter and identify information relevant to users in specific fields, without requiring excessive time and effort. The processes of searching and analyzing information can be exhausting, often involving the review of large volumes of text and diverse sources. This paper presents the application of a multi-agent system for the automated collection, filtering, and analysis of information, using the CrewAI framework, with a focus on the cryptocurrency market. The system enables users to efficiently access concise and relevant insights on cryptocurrency price movements and market trends with minimal manual input. It is powered by a semantic search engine that allows meaning-based retrieval of content, ensuring access to highly relevant data from a variety of sources. This approach supports effective market analysis and trend forecasting, ultimately helping economists and investors make better decisions, optimize resources, and reduce risk.

Abstract: The exchange of Electronic Health Records (EHR) with research organizations is vital for advancing scientific discoveries and improving patient care. However, healthcare providers face labor-intensive approval processes, privacy risks, and a lack of interoperability across EHR formats. In this paper, we propose a patient-centric approach for automated medical record exchange that ensures secure data transfer, maintains patient anonymity, and reduces administrative overhead, thereby enabling efficient data sharing across institutions. We implement this approach through a system that leverages blockchain technology with smart contracts for immutable tracking of user actions and automated access permission handling, while specialized database ensures real-time performance, scalability, and flexible off-chain storage for diverse EHR formats.

Abstract: As privacy regulations and security concerns increase, database encryption techniques have become essential in modern applications. Multiple approaches exist for querying encrypted data, including Order-Preserving Encryption, Fully Homomorphic Encryption, Oblivious RAM, and Trusted Execution Environments. However, these come with serious security or performance drawbacks, making them impractical for majority of real-world applications. This paper focuses on two widely used encryption strategies: Searchable Encryption and Structured Encryption. Searchable Encryption enables exact-match searches on encrypted fields, whereas Structured Encryption encrypts entire structured objects but requires full decryption on retrieval. Both approaches align with Endto-End Encryption principles, ensuring that sensitive data remains encrypted throughout its lifecycle, with decryption keys held only by the data owner. We present a comparative study evaluating the impact of Searchable Encryption and Structured Encryption on query performance, storage overhead, and computational cost in a relational database setting. The results of our experiment provide insights into the efficiency of encrypted query processing and are aimed to help developers in selecting the most suitable encryption method for secure data storage.

Abstract: Learning word embeddings on large, unlabeled corpora has proven effective for many natural language tasks. However, these representations can be further improved by incorporating external lexical resources. Previous research has demonstrated that lexical vector representation (embeddings, e.g., Dict2vec) trained on both text and lexical data (e.g., WordNet and/or monolingual dictionaries) gives improved results for English. The existence of the Serbian Wordnet and several Serbian electronic dictionaries enables testing this approach for Serbian within this project. In this paper, we adapt the original Dict2vec project for Serbian language resources. We present the textual, lexical, and vector resources prepared and used for training and evaluation, describe the training pipeline, and discuss preliminary evaluation results. We conclude this paper by outlining ongoing work and future steps.