Švarcmajer, Miljenko; Ivanović, Denis; Rudec, Tomislav; Lukić, Ivica,
Technologies 2025, 13(1), DOI: 10.3390/technologies13010033
Abstract:
This paper investigates the application of graph theory and variants of greedy graph coloring algorithms for the optimization of distributed peer-to-peer networks, with a special focus on private blockchain networks. The graph coloring problem, as an NP-hard problem, presents a challenge in determining the minimum number of colors needed to efficiently allocate resources within the network. The paper deals with the influence of different graph density, i.e., the number of links, on the efficiency of greedy algorithms such as DSATUR, Descending, and Ascending. Experimental results show that increasing the number of links in the network contributes to a more uniform distribution of colors and increases the resistance of the network, whereby the DSATUR algorithm achieves the most uniform color saturation. The optimal configuration for a 100-node network has been identified at around 2000 to 2500 links, which achieves stability without excessive redundancy. These results are applied in the context of a private blockchain network that uses optimal connectivity to achieve high resilience and efficient resource allocation. The research findings suggest that adapting network configuration using greedy algorithms can contribute to the optimization of distributed systems, making them more stable and resilient to loads.
Keywords:
greedy algorithms; graph coloring; DSATUR; distributed systems; peer-to-peer networks; connectivity optimization; private blockchain network; graph theory
Krčmar, Tea, Šabanović, Dina; Habijan, Marija; Galić, Irena; Lukić, Ivica
Abstract:
Efficient image sampling is essential for balancing acquisition speed, resolution, and computational cost in computer vision. Insufficiently sampled data often leads to quality degradation and reconstruction artifacts, posing challenges for downstream applications. To address this, we developed a diffusion-based model with a dynamic adaptive sampling strategy, combining randomly generated masks with denoising to enhance reconstruction quality. Comprehensive analyses validate the model’s effectiveness, highlighting its potential to optimize workflows in data-intensive image processing applications.
Lukić, Ivica; Kohler, Mirko; Krpić, Zdravko; Švarcmajer, Miljenko
Technologies 2025, 13(7), 300; https://doi.org/10.3390/technologies13070300
Abstract:
This paper presents an integrated Smart City platform that combines digital twin technology, advanced machine learning, and a private blockchain network to enhance data-driven decision making and operational efficiency in both public enterprises and small and medium-sized enterprises (SMEs). The proposed cloud-based business intelligence model automates Extract, Transform, Load (ETL) processes, enables real-time analytics, and secures data integrity and transparency through blockchain-enabled audit trails. By implementing the proposed solution, Smart City and public service providers can significantly improve operational efficiency, including a 15% reduction in costs and a 12% decrease in fuel consumption for waste management, as well as increased citizen engagement and transparency in Smart City governance. The digital twin component facilitated scenario simulations and proactive resource management, while the participatory governance module empowered citizens through transparent, immutable records of proposals and voting. This study also discusses technical, organizational, and regulatory challenges, such as data integration, scalability, and privacy compliance. The results indicate that the proposed approach offers a scalable and sustainable model for Smart City transformation, fostering citizen trust, regulatory compliance, and measurable environmental and social benefits.
Keywords:
digital twin; blockchain; business intelligence; smart city; ETL; machine learning; participatory governance; urban sustainability
Miljenko Švarcmajer, Tea Krčmar, Dina Šabanović, Ivica Lukić
ELMAR 2025, DOI: 10.1109/ELMA66948.2025.11193746
Abstract:
The growing reliance on cloud-based business intelligence systems has increased the demand for secure communication and verifiable data integrity. However, the emergence of quantum computing poses a significant threat to current cryptographic protocols, as widely adopted encryption schemes may become vulnerable to quantum-enabled attacks. In this paper, the vulnerabilities introduced by quantum advancements in contemporary business intelligence systems are examined. To address these challenges, the integration of quantum-resistant cryptographic components—such as post-quantum digital signatures and PQ-DPoL consensus mechanisms—is proposed. A security model based on permissioned blockchain architecture is presented as a practical and scalable solution for long-term resilience, offering enhanced protection in anticipation of future quantum capabilities.
Krpić, Zdravko; Lukić, Ivica; Habjan, Marija; Loina, Luka
Future Generation Computer Systems, DOI: 10.1016/j.future.2025.108137DOI
Abstract:
Single Board Computer Clusters (SBCCs) are increasingly used as accessible, low-power platforms for parallel and distributed computing, particularly in edge and fog environments. Yet their performance remains underexplored through reproducible, tuned evaluations. This paper presents a benchmarking methodology based on the High Performance Linpack (HPL) benchmark, selected for its use of dense linear algebra kernels common in scientific and machine learning workloads. The evaluation includes HPL parameter tuning, compiler configuration, and comparison of ATLAS vs. OpenBLAS.
We apply the methodology SBCs spanning a decade of development: Raspberry Pi 1B, 3B, 4B, and 5, Cubieboard 2, Odroid U3, and Odroid-MC1. Results show that software-level tuning without overclocking or hardware modification can yield performance improvements of up to 2.3*over prior reports. A 146*increase in HPL performance between the Pi 1B and Pi 5 illustrates the evolution in computational capability within a stable form factor. OpenBLAS outperforms ATLAS on newer platforms, while ATLAS retains marginal advantages on older boards.
The findings provide a reproducible baseline for SBCC performance evaluation and support their relevance for benchmarking, education, and energy-efficient high-performance workloads in scenarios where conventional clusters are impractical due to cost, size, or power.
Ivica Lukić, Tea Krčmar, Dina Šabanović, Miljenko Švarcmajer, Mirko Köhler, Marko Carević
Abstract:
Unstructured data challenges traditional business intelligence (BI), particularly when processing layout-rich documents like scanned images or PDFs. This paper explores using large language models (LLMs) for extracting structured information, focusing on the LMDX (Language Model-based Document Information Extraction and Localization) approach. LMDX encodes document layout by inserting coordinate tokens directly into the prompt, eliminating the need for vision-based models or architectural changes. We also propose a blockchainenabled, cloud-based pipeline that integrates LMDX for secure and scalable document intelligence. The paper presents key findings and outlines future research to enhance document processing and automated decision-making in modern BI systems.
Krčmar, Tea; Šabanović, Dina; Švarcmajer, Miljenko; Lukić, Ivica,
Mach. Learn. Knowl. Extr. 2026, 8(3), 60; https://doi.org/10.3390/make8030060
Abstract:
Anomaly detection on tabular data is widely used in fraud detection, predictive maintenance, and medical screening. While heterogeneous ensembles combining multiple detection paradigms achieve strong performance, their computational cost limits deployment in latency-sensitive or resource-constrained environments. We propose KD-AnomalyNet, a teacher–student framework that distills anomaly knowledge from a high-capacity ensemble into a lightweight neural model for efficient inference. Beyond performance replication, we study how anomaly representations transfer during distillation. To this end, we introduce a noise perturbation analysis that serves as a diagnostic probe for representation stability without introducing additional trainable components. Experiments on ten benchmark datasets show that the distilled model preserves up to 98.5% of the teacher’s AUC-ROC on the nine capacity-sufficient datasets (84.7% mean retention across all ten datasets) while achieving 26–181× inference speedups. Our analysis reveals which forms of anomaly knowledge transfer reliably—global outliers (78% transfer) and isolation-based detection (88% retention)—and which degrade under compression—local outliers (20% transfer) and neighborhood-based detection (76% retention)—providing practical guidance for deploying distilled anomaly detectors.
Keywords:
anomaly detection; knowledge distillation; teacher–student learning; ensemble methods; tabular data; temperature scheduling; lightweight models