ABSTRACT. In the landscape of black-box optimisation, benchmarking plays a crucial role in understanding algorithm behaviour and guiding algorithm design. Yet, much of current practice remains focused on surface-level comparisons, offering limited insight into why certain algorithms perform well and how specific components contribute to their success or failure.

This talk explores a new direction: explainable benchmarking. Drawing on techniques from explainable AI, it becomes possible to move beyond performance tables, towards uncovering the underlying mechanisms that drive algorithmic behaviour. The IOHxplainer framework demonstrates how performance data—collected across a broad range of configurations and problem instances—can be analysed to attribute outcomes to individual components and hyperparameters, especially for highly modular and parameterised algorithm designs. Examples from recent studies on modular CMA-ES illustrate how these methods reveal evolving module contributions over time, quantify sensitivity to random seeds and problem instances and identify design choices that genuinely matter. Being robust and scalable, this approach opens up new possibilities for principled algorithm development, configuration and selection.

The talk concludes with reflections on the future of benchmarking: how to make it more representative, more insightful and ultimately more useful for building optimisation methods that are not only powerful but also better understood.

ABSTRACT. In this preliminary contribution we study the \emph{kernel trick} under the eyes of the Hilbert space instead of the kernel itself. This leads to construction of reproducing kernels (RKs) associated to a given Hilbert space. We show that a large family of classical Sobolev spaces where the ``trace'' map is continuous are reproducing kernels Hilbert spaces and deduce a universal reproducing kernel for $H^{m+1}(\mathbb R^{2m})$ as well as for $H^{m+1}(\mathbb R^{2m+1})$. For $\Omega\subset\mathbb{R}^2$ bounded, we construct RKs for (a subspace of) $H^2(\Omega)$ using finite elements. These reproducing kernels may lead to new nonlinear separators for support vector machine linked only to the Solobev regularity of the separator.

ABSTRACT. Many ways of solving the popular Sudoku puzzle exist and while solving the puzzle itself is an NP-complete task, some methods manage to solve the 9x9 variant in a reasonable amount of time. However, there have been very few studies that directly compare these methods while providing concrete performance metrics. This paper aims to deliver statistically relevant key performance indicators for a variety of the major techniques used to solve the Sudoku puzzle.

Moreover, the obtained results take into consideration variables such as: standard deviation, median and mean that are calculated using statistical methods % that are recognized in the scientific literature like: warm-up rounds, bootstrap resampling, repeated measurements.

The study proves that certain algorithms perform well all-around with no changes in performance with regards to different difficulties, but some degrade in performance when solving more difficult variants of the puzzle. The popular Constraint Satisfaction Problem (CSP) encoding of Sudoku suffers from severe performance degradation as the difficulty range of the puzzles increases with an observed decrease of up to 200\% in performance. This performance degradation also affects the Dancing Links (DLX) approach, but the performance degradation slope is much leaner than the CSP one. DLX records a maximum degradation of 24.78\% in solving the hardest rated puzzles.

However, the Satisfiability (SAT) method registers no performance loss when increasing the difficulty of the Sudoku puzzles.

This paper positions itself as a novel analysis that takes into account many technical and statistical aspects.

ABSTRACT. In this paper, a new adaptive neural network is proposed for a class of uncertain nonlinear systems with unknown control direction. The control design can be applied to deal with a class of nonlinear systems without the need to satisfy the Lipchitz condition, where the RBF neural networks are employed to approximate the unknown nonlinear functions. Unlike previous iterative learning control (ILC) schemes, the Nussbaum function method is not utilized to cope with the unknown control direction. In fact, in this paper, the unknown input control is adjusted using an adaptation law. Another advantage of the control is that there is no constraint on nonlinearities. The stability of the closed loop learning system is guaranteed. This proof is based upon the use of a Lyapunov-like positive definite sequence, which is shown to be monotonically decreasing under the proposed control scheme. Finally, an illustrative example based on inverted pendulum systems is provided to demonstrate the effectiveness of the proposed controller.

ABSTRACT. In this paper we propose an improved model for multidimensional data representation that is based on the PCA technique for optimal reduction of the input data space. The presented model carries out a study on the optimal value of the size of the projection space of multidimensional image data so as to obtain improved results at the level of correctness percentages in the problem of classifying bidimensional signals. To determine the projection parameter, a search scheme based on the nearest neighbors algorithm (KNN) together with an evolutionary algorithm is used. To establish the projection model of the input data in the process of detecting the appropriate values to provide the best possible results in the classification of new signals, several measures of data similarity are used to determine the nearest neighbors. In the last part of the paper, the results of the experimental study regarding the performance of the optimal reduction model of the processed multidimensional data space are presented.

ABSTRACT. Large language models often excel in fluency but lack alignment with human values and transparency in their reasoning. We introduce X-RLHF, a framework that augments Reinforcement Learning from Human Feedback (RLHF) with internal interpretability metrics (Integrated Gradients and attention) via a composite reward function. Fine-tuning Falcon3-1B reveals that internal signals enhance reasoning coherence without harming output quality, whereas the external reward head struggles to mirror human preferences. X-RLHF demonstrates that embedding explainability objectives into RLHF yields more transparent, trustworthy models. We conclude by outlining challenges and future directions for scaling and richer interpretability methods.

ABSTRACT. In this work, we present a novel approach for Explainable Artificial Intelligence (XAI) by employing computational topology techniques as feature attribution methods. By converting density gaps into one-dimensional persistence scores, we obtain a model-agnostic measure of each feature’s intrinsic importance. Tested on the Breast Cancer Wisconsin benchmark dataset, along a similar synthetic dataset with a known ground truth, our scores based on topology recover true feature ranks nearly as well as SHAP (Spearman $\rho\approx0.6$) and show agreement with the ground truth ($\rho\approx0.5$). These results demonstrate that coupling topology with attribution yields concise explanations that enhance trust and reveal hidden feature interactions, pushing XAI beyond purely model-centric viewpoints.

ABSTRACT. The growing accessibility of artificial intelligence has led to widespread experimentation with deep learning models. As neural networks become increasingly common, efforts to optimize the training process have intensified, focusing on architectural improvements, better optimization algorithms, and learning rate scheduling techniques. Batch size schedulers have recently gained attention for their potential to improve convergence speed and hardware efficiency, analogous to the benefits provided by learning rate schedulers.

Increasing the batch size during training enables efficient hardware utilization. However, larger batch sizes are often associated with reduced generalization performance. To address this issue, adaptive strategies have been proposed that incrementally adjust batch size during training. These approaches aim to maintain model performance while benefiting from faster convergence and improved compute efficiency.

In this work, we analyze the performance implications of using batch size schedulers, both in terms of training time and generalization ability. We compare six different batch size schedulers against equivalent learning rate adaptation policies on image classification benchmarks. Our results show that batch size schedulers can reduce training time by up to 38% with a generalization drop of less than 0.7%.

Finally, we provide practical recommendations for integrating batch size schedulers into modern training pipelines and replacing learning rate schedulers.

ABSTRACT. Teaching abstract calculus concepts, such as limits of functions, remains a challenge in high school education, often failing to accommodate diverse learning paces. This study explores the design, implementation, and evaluation of an AI-powered mobile application that offers personalized feedback and gamified learning experiences to enhance student engagement and understanding. The application integrates AI-based mistake analysis, dynamic exercise generation, and a reward system. It was tested through a user study where students used the application and provided structured feedback. Results show that learners highly valued the AI explanations and gamification features. The findings suggest that combining AI-driven personalization with gamification can significantly improve comprehension and engagement in STEM education, highlighting the potential of such tools in modern learning environments.

ABSTRACT. In this work, we present YMCQ, our novel method for automated multiple-choice question (MCQ) generation that enhances both answer quality and distractor plausibility through reasoning-based explanations. Our approach addresses well-known challenges in MCQ generation, including the lack of diversity and plausibility in distractors, as well as the overreliance on closed-source models. To overcome these limitations, we leverage open-source large language models (LLMs) and enrich standard MCQs with explicit reasoning processes behind both correct answers and distractors, including misconceptions. By doing so, we aim to bring deeper educational value to automatically generated questions while keeping our solution cost-effective and widely accessible.

ABSTRACT. This article explores how reaction-diffusion models, which traditionally describe chemical and biological pattern formation, can be applied to dynamic clustering problems. We present an overview of the underlying mathematical framework for reaction-diffusion systems, discuss the mechanisms by which they can effectively discover natural groupings. The self-organizing properties intrinsic to reaction-diffusion systems offer new computational paradigms for tackling large-scale clustering challenges, potentially leading to more efficient and adaptive algorithms.

ABSTRACT. Human Action Recognition plays a vital role in assistive technologies deployed in real-world scenarios, where models must address challenges related to privacy, contextual reasoning, and robustness beyond controlled benchmarks. In this work, we introduce a novel taxonomy that classifies HAR models along two primary axes: the level of privacy ensured by the input modality and the model’s readiness for deployment in unconstrained environments. This classification is further informed by two semantic dimensions -- temporal modeling and contextual awareness, which capture the model's ability to reason over time and integrate environmental cues. Based on this taxonomy, we define the Privacy-Conscious Readiness and Modality Semantics score, a composite metric prioritising privacy while rewarding semantic richness and deployment feasibility. Our contribution enables a structured and ethics-aware evaluation of HAR models, facilitating their responsible use in assistive applications.

ABSTRACT. The continuous functioning of any industrial manufacturing facility, especially critical infrastructures, became crucial in the current multi risk context. Monitoring and detection of anomalies carries multiple significant practical benefits that are direct Industry 4.0 goals, and some of them improving resiliency and sustainability, implicitly targets of Industry 5.0. For this reason, the current paper explores the usage of feedforward autoencoder neural networks for anomaly detection. The proposed approach is designed to capture deviations in the overall operational behavior of a plant, enabling system-wide monitoring rather than being constrained to the identification of specific, predefined fault scenarios. The obtained autoencoder was subject to further experimental testing on synthetic data and a direct comparison with five other anomaly detection methods (Z-score, Interquartile Range, Isolation Forest, One-Class Support Vector Machines and Local Outlier Factor) proves superior performance from the autoencoder in terms of precision, recall, and F1 score. The foreseen case study was focused on data from a real drinking water treatment plant.

ABSTRACT. In today’s landscape, where cyber threats evolve on all fronts, the human factor remains a weak link that may contribute to the success of cyberattacks. This paper presents a comprehensive solution for psychological assessment and cybersecurity risk management by profiling individuals within an organization and estimating the likelihood that a phishing, ransomware, or social-engineering attack against them will succeed and propagate through the infrastructure. Our architecture is built as a set of microservices with distinct roles. One component processes incoming text and extracts five relevant psychometric dimensions - awareness, conscientiousness, stress, neuroticism, and risk tolerance - using a BERT-based fine-tuned model and heuristic relabeling rules. Another component functions like an HR department, storing and relating users and organizational entities in a relational database. The third calculates each person’s composite risk via advanced techniques, uncovers vulnerabilities in a simulated cyberattack scenario, and parses BPMN diagrams for interoperability. All services are exposed through a graphical interface designed to deliver high-performance support for organizational cybersecurity risk management, featuring real-time simulations, scenario customization, and an intuitive user experience. Experimental evaluations demonstrate superior performance in detecting attack susceptibility and provide a replicable framework for integrating the human factor into modern cybersecurity risk strategies.

ABSTRACT. This survey investigates the application of Facial Action Units (FAUs) in detecting deepfake videos, focusing on both technical and algorithmic dimensions. By analyzing micro-expressions such as eye gaze, lip synchronization, and head poses - often overlooked by generative models - the study evaluates the viability of FAUs as biometric indicators for manipulation. Using OpenFace for facial feature extraction, we review both machine learning and numerical approaches, and compare detection techniques based on their accuracy, efficiency, and resilience to tampering. We also assess major public datasets (e.g., FF++, DFDC, Celeb-DF) to understand limitations in generalization and propose future directions for scalable, privacy-preserving FAU-based detection systems. The findings underline the potential of combining spatiotemporal modeling with FAU dynamics to achieve robust and interpretable deepfake detection.

ABSTRACT. The large spread of disinformation across digital platforms creates significant challenges to information integrity. This paper presents a multi-agent system that uses relation extraction to detect disinformation in news articles, focusing on titles and short text snippets. The proposed Agentic AI system combines four agents: (i) a machine learning agent (logistic regression), (ii) a Wikipedia knowledge check agent (which relies on named entity recognition), (iii) a coherence detection agent (using LLM prompt engineering), and (iv) a web-scraped data analyzer that extracts relational triplets for fact checking. The system is orchestrated via the Model Context Protocol (MCP), offering shared context and live learning across components. Results demonstrate that the multi-agent ensemble achieves 95.3\% accuracy with an F1 score of 0.964, significantly outperforming individual agents and traditional approaches. The weighted aggregation method, mathematically derived from individual agent misclassification rates, proves superior to algorithmic threshold optimization. The modular architecture makes the system easily scalable, while also maintaining details of the decision processes.

ABSTRACT. In this work, we compare three mixed-integer programming formulations for the Bandwidth Coloring problem: (i) the Assignment-based model, (ii) the Hybrid Partial-Order model, and (iii) the Big-M model. We perform computational experiments on benchmark instances. The numerical results demonstrate that the Big-M model consistently provides better performances in terms of computational time. The empirical evidence suggests that, under standard branch-and-cut settings, the Big-M model is the most effective among the three when tackling instances of medium to large size.

ABSTRACT. Maintaining high-quality code and efficiently managing software issues are critical aspects of modern software development. As software projects grow in complexity, developers face increasing challenges in handling technical debt, prioritizing tasks, and addressing issues effectively. To tackle these challenges, this project presents multiple AI-based models designed to classify SonarQube issues into fault-prone and non-fault-prone categories and to prioritize GitHub issues into three urgency levels: High, Medium, and Low priority. By leveraging machine learning models, the system automates these processes, significantly reducing manual workload while ensuring that the most critical issues receive timely attention.

This project integrates data from both SonarQube and GitHub APIs, applying advanced classification and prioritization techniques. This approach not only enhances issue management but also contributes to improving overall software quality by helping developers focus on the most impactful problems first.

ABSTRACT. Detecting fraudulent transactions in financial systems requires models that balance accuracy with interpretability. Although black-box machine learning models offer high predictive performance, their opacity limits trust and practical deployment in high-stakes domains like fraud detection. In this work, we present a Genetic Programming (GP) framework that evolves interpretable Boolean rule-based classifiers tailored for fraud detection. Our approach leverages both mutation and crossover operations to explore the space of candidate rules, enabling global optimization, and addresses the limitations of local search strategies. To ensure interpretability, we transform numerical features using a decision tree-based binarization technique that extracts class-aware binary thresholds. This allows the GP to evolve rules composed entirely of logical operators over binarized features, creating clear and interpretable models. Evaluated on the PaySim dataset, our approach exceeds the performance of other interpretable methods and remains competitive when compared with the state-of-the-art black-box models. Our results show that GP-evolved Boolean rules can serve as interpretable alternatives for real-world fraud detection systems.

ABSTRACT. The stock market has a crucial role in the economy of a country. While anomalies in the stock market cannot cause crashes, they foster ideal conditions that lead to them. It is essential to investigate such anomalies, in order to have a stable economy. In the literature clustering algorithms seem to dominate as the primary method for detecting anomalies in the stock market; however, other studies have proven that autoencoders perform better at discovering anomalies in time-series data. We propose a novel approach that combines autoencoders with a post-training genetic algorithm to optimize the latent space, thus improving the accuracy overall. Unlike conventional approaches that use genetic algorithms for training or hyperparameter tuning, our method refines the detection process by improving the accuracy of the reconstruction. On a real-world financial dataset, our method achieved an AUC of 0.9515, outperforming both traditional clustering models and baseline autoencoders.

ABSTRACT. This paper was submitted for review to a different journal. High-resolution synthetic aperture radar imagery is essential in mining applications, particularly for early hazard detection and ground deformation assessment. Nonetheless, missions like COSMO-SkyMed have a more restricted spatial and temporal scope. Simultaneously, the publicly accessible Sentinel-1 exhibits inferior spatial resolution and lacks of phase information, which is useful for InSAR analysis. To address these limitations, we propose a deep learning architecture tailored for cross-sensor SAR super-resolution, capable of converting Sentinel-1 ground-range intensity data into high-resolution, complex-valued COSMO-SkyMed imagery. The framework, collaboratively trained with a hybrid loss that integrates both amplitude and phase fidelity, consists of a U-Net-based cross-sensor translation network (XSARNet) and a complex-valued super-resolution network (DC2SCN4x). In order for it to be compliant with downstream InSAR, the model retains real and imaginary components while overseeing deformation actions. We compile a multi-sensor, multi-temporal dataset for training purposes in the Gummern mining district of Austria. We assess the generality in the Finnish mining region utilising this dataset. This technique enhances spatial resolution, leading to improved results on SAR-specific and perceptual measures. Both quantitative and qualitative evaluations indicate that it also maintains complicated coherence. This method facilitates the monitoring of mining-related dangers that are scalable and economical, utilising readily available Sentinel-1 data.

ABSTRACT. Landslides pose dangerous risks to human safety and to infrastructure, especially in densely populated or geologically unpredictable regions. Interferometric Synthetic Aperture Radar (InSAR) technology, along with time series analysis has significantly improved the ability to monitor ground deformations associated with such hazards. This study explores the detection of change points in InSAR-derived ground motion time series from the Praid region in Romania, based on Sentinel-1 data collected over a period of five years. The methodology starts with clustering Persistent Scatterers (PS) with similar deformation patterns using the HDBSCAN algorithm, followed by the application of several detrending methods to enhance change point detection accuracy. Four algorithms, Pruned Exact Linear Time (PELT), Binary Segmentation (BS), Window Sliding and Bottom-Up Segmentation, are implemented to identify abrupt fluctuations in ground motion. The absence of ground truth data motivated the selection of a consensus-based on majority voting as evaluation system. More specifically, the validation of a detected change point was correlated with the number of algorithms that identified it independently, with a minimum of two algorithms out of the four implemented. The findings offer insights into deformation activity and highlight the value of change point analysis in landslide detection. Future work will include a more comprehensive analysis of data particularities, the integration of contextual environmental factors and finally, the application of advanced time series techniques, to further enhance detection accuracy and improve early warning and risk management.

ABSTRACT. Desertified areas represent a growing environmental challenge in Romania, particularly in the southern regions, where soil degradation is exacerbated by climate change and unsustainable land use. Despite the increasing severity of this issue, the lack of efficient and scalable monitoring techniques based on satellite imagery hampers accurate assessment and timely intervention.

This study explores the spatial and temporal dynamics of desertification in southern Romania by leveraging multispectral imagery and autoregressive forecasting techniques. Several key spectral indices, i.e., the Bare Soil Index, Normalized Sand Index, Normalized Difference Enhanced Sand Index, and the Crust Index (which is sensitive to soil sealing and biological crusts) were employed to highlight vulnerable surfaces. Various autoregressive methods such as ARIMA, SARIMA and AutoARIMA were used in combination with several adaptive image thresholding techniques to forecast the evolution of desertified areas in a region from southern Romania.

The results reveal a notable expansion of sandy and barren land strongly correlated with vegetation loss and increased surface dryness due to prolonged droughts. Forecasts from the tested models indicate an increase that will last at least until 2030 but whose prediction is highly influenced by the used thresholding method.

ABSTRACT. Traditional GIS systems have limited capabilities in handling dynamic and multidimensional geospatial data. While MultiGIS extends GIS capabilities by integrating cross-domain and real-time data, its implementation faces key technical, data-related, and social challenges, including handling real-time heterogeneous data, insufficient semantic interoperability, and ethical risks related to privacy, equity, and data bias. This paper proposes AI4MultiGIS, a unified modular framework that aims to overcome these limitations through five interoperable layers: data perception, semantic fusion, artificial intelligence, decision reasoning, and visual interaction. Its effectiveness has been validated through three pilot cases: SuDS planning in the UK, invasive species monitoring in Romania, and heritage analysis in Saudi Arabia, demonstrating the framework's adaptability, interpretability, generality, and ethical robustness in different application scenarios.

ABSTRACT. Convolutional neural networks (CNN) have shown great results for object-detection tasks by learning texture and pattern-extraction filters. However, object-level interactions are harder to grasp without increasing the complexity of the architectures. On the other hand, Point Process models propose to solve the detection of the configuration of objects as a whole, allowing the factoring in of the image data and the objects prior interactions. In this talk, we propose combining the information extracted by a CNN with priors on objects within a Markov Marked Point Process framework. We also propose a method to learn the parameters of this Energy-Based Model. We apply this model to the detection of small vehicles in optical satellite imagery, where the image information needs to be complemented with object interaction priors because of noise and small object sizes. This is a joint work with my former PhD student (Jules Mabon) in collaboration with Airbus DS (Mathias Ortner).

ABSTRACT. Interference is a phenomenon on proof systems for SAT solving that is both counter-intuitive and bothersome when developing proof-logging techniques. However, all existing proof systems that can produce short proofs for all inprocessing techniques deployed by SAT present this feature. Based on insights from propositional dynamic logic, we propose a framework that eliminates interference while preserving the same expressive power of interference-based proofs. Furthermore, we propose a first building blocks towards RUP-like decision procedures for our dynamic logic-based frameworks, which are essential to developing effective proof checking methods.

ABSTRACT. To solve classification problems, we present a novel SAT-based framework for learning NNF networks directly from binary input-output data. In analogy to deep neural networks, negation normal form (NNF) networks exhibit a deep structure with learnable Boolean weights. By encoding this learning prob- lem as a propositional satisfiability instance, our method leverages modern SAT solvers to construct logically correct, compact, and interpretable Boolean models of given datasets. Unlike statistical learners or heuristic rule induction algorithms, our approach guarantees logical fidelity on the training data. Evaluations on synthetic benchmarks and real-world datasets demonstrate that it outperforms state-of-the-art rule-based algorithms as well as a greedy NNF learning algorithm in terms of accuracy.

ABSTRACT. During the last years, real-time applications have increasingly employed multicore architectures, in response to the continuous growth in the complexity of the problems they are approaching. A large variety of structures have been proposed, leading to significantly different approaches. NUMA (Non-Uniform Memory Access) architectures are widely used, which comes as no surprise since, from a historical point of view, many embedded systems have been using local memories. While the use of local memories is key to increasing performance, the asymmetry of the access times raises difficult questions regarding the efficient management of the available resources, especially when shared variables are present. The issue is even more obvious in automotive systems, where the large amount of information involved pushes the size of real-life problems to the limits of intractability and thus requires innovative approaches. A recent example is [1], which is approaching the allocation of the jobs’ data and code to the processors in a NUMA system. As the problem is intractable for common automotive applications, the idea is to decompose it into smaller subproblems, to be dealt with sequentially. Nevertheless, because multiple phases rely on Integer Linear Programming, tractability is still a matter of concern. This paper proposes an alternative to the existing solution, based on the same principles. A rethinking of the decomposition into subproblems allows a simpler (non-ILP) solution to one phase, while for another one the ILP is simplified. Experimental results show that, while the new approach is not necessarily more efficient, it is comparable to the existing one, to a degree that justifies the gain in simplicity.

ABSTRACT. Codifying mathematical theories in a proof assistant or computer algebra system is a challenging task, of which the most difficult part is, counterintuitively, structuring definitions. This results in a steep learning curve for new users and slow progress in formalizing even undergraduate level mathematics. There are many considerations one has to make, such as level of generality, readability, and ease of use in the type system, and there are typically multiple equivalent or related definitions from which to choose. Often, a definition that is ultimately selected for formalization is settled on after a lengthy trial and error process. This process involves testing potential definitions for usability by formalizing standard theorems about them, and weeding out the definitions that are unwieldy.

Inclusion of a formal definition in a centralized community-run mathematical library is typically an indication that the definition is “good.” For this reason, in this survey, we make some observations about what makes a definition “good,” and examine several case studies of the refining process for definitions that have ultimately been added to the Lean Theorem Prover community-run mathematical library, mathlib. We observe that some of the difficulties are shared with the design of libraries for computer algebra systems, and give examples of related issues in that context.

ABSTRACT. The presented paper discusses a hybrid algorithm for minimizing the power consumed by the finite state machine (FSM). The proposed algorithm is based on the modification of three algorithms: the sequential algorithm, where code assignment is based on the previous encoding, state splitting, where a state is divided into several states, and using a special model of an FSM, which lead to the reduction of the length of the internal state's code. Modifications to the algorithms include using the second or third best element in the sequential algorithm, splitting the state into more than two new states, and using different lengths for state code. Every modification is tested to see if it leads to a decrease in the power in the FSM. The proposed algorithm executes methods starting from the state splitting, followed by the implementation of the common architectural model, and finally, the sequential method of state encoding. The experimental results show that the proposed method reduces the power consumption compared to NOVA, JEDI, column-based based and original sequential algorithms by at least 16\%.

ABSTRACT. Control-Flow Graphs (CFGs) are fundamental structures in the static analysis of binary executables, providing insights into program behavior, logic, and control dependencies. In the context of malware analysis, CFGs serve as critical tools for reverse engineering, behavioral profiling, and the detection of obfuscation techniques. This study presents a systematic approach for the automated generation and extraction of Control-Flow Graphs from malicious binaries using the Ghidra reverse engineering framework and its scripting API. By leveraging Ghidra's decompilation and symbol recovery capabilities, the proposed method enables efficient identification of function boundaries, control structures, and execution paths within obfuscated and packed binaries. The implementation facilitates scalable and repeatable CFG extraction for large malware datasets, supporting downstream tasks such as classification, similarity analysis, and feature engineering for machine learning pipelines. Experimental validation on multiple malware samples demonstrates the effectiveness and robustness of the approach in recovering meaningful control-flow representations despite the presence of common evasion techniques.

ABSTRACT. Reliable state-of-health (SOH) prediction is critical for ensuring the safety and dependability of batteries. Traditional data-driven techniques primarily rely on historical battery data to forecast future performance. However, since the degradation of the battery SOH is significantly influenced by upcoming operational loads, the use of historical information alone can limit the accuracy of the prediction. To address this, a novel prediction framework is proposed that incorporates different battery factors to predict SOH. The suggested approach utilizes an XGBoost classification model tuned by a modified version of the particle swarm optimization algorithm to perform the battery health prediction task. The best models produced promising outcomes with an accuracy of approximately 87.17%, showing prospect in possible applications in this field.

ABSTRACT. The Internet of Things~(IoT) plays an ever increasingly important role in everyday life. Devices with constant internet connectivity are now common in homes, industrial and other professional environments, as well as part of smart energy grids and management systems. However, given the often limited computational resources of these devices, security is often given lower priority. As IoT devices can access the Internet, they can be used as parts of botnets and used to execute various cyberattacks on the network. Such vulnerabilities are especially crucial when they concern critical infrastructure and data. This study evaluates the effectiveness of a machine learning (ML) based solution for IoT threat detection. Given the dynamic nature of IoT networks, and the continuous development of new vectors of attack, an adaptive approach needs to be taken to maintain security. A modified metaheuristic optimizer is presented and used to tune classifier hyperparameters to ensure favorable performance. The solution being introduced has been applied to authentic data and the simulations suggest promising outcomes. The best profiteering models reach an accuracy rating of 0.99361.

ABSTRACT. Scheduling in dynamic environments represents a complex challenge in the field of combinatorial optimization, with applications across various industries, including manufacturing and logistics. Unlike static scheduling, which assumes complete prior knowledge of events and constraints, dynamic environments involve uncertainties such as unexpected task arrivals, delays, and resource modifications, all requiring rapid plan adjustments. This paper explores the use of Genetic Programming (GP) as a method for addressing dynamic scheduling problems. Genetic Programming provides a flexible and adaptive framework for generating scheduling rules, optimizing resource allocation, and task sequencing under uncertain conditions. The model considers diverse constraints, such as precedence constraints between tasks belonging to different jobs, available resources, and adherence to deadlines, aiming to create robust and efficient solutions. To evaluate the methodology's performance, a set of test problems was generated to simulate unexpected events and unpredictable variables, enabling the testing of GP's ability to react to real-time changes. The study contributes to the application of GP in dynamic scheduling, demonstrating its potential to solve complex problems and laying the groundwork for advanced methods in real-world applications.

ABSTRACT. Hyperparameter optimization (HPO) represents a computationally demanding yet crucial phase in developing machine learning models. Differential Evolution Hyperband (DEHB) combines evolutionary search with efficient resource allocation to address this challenge. A key implementation detail in DEHB involves handling boundary constraint violations when mutation and crossover operations generate infeasible solutions. This study systematically evaluates nine boundary constraint handling methods (BCHMs) within DEHB for optimizing neural network hyperparameters across five classification benchmarks. The experimental analysis reveals that methods incorporating information from the best-found solution consistently outperform simpler approaches, achieving superior validation loss and accuracy. Surprisingly, strategies with lower repair frequencies do not necessarily yield better optimization performance, suggesting that controlled boundary violations followed by intelligent correction can enhance exploration. These findings demonstrate that BCHM selection significantly impacts DEHB's effectiveness and provide practical guidance for practitioners implementing multi-fidelity hyperparameter optimization.

ABSTRACT. Congenital heart anomalies are among the most frequent malformations in newborns, emphasizing the importance of early diagnosis. However, identifying heart anomalies in the first trimester is particularly challenging due to the incomplete development of fetal cardiac structures. To support medical specialists during this difficult stage, this study proposes an artificial intelligence system to analyze first-trimester ultrasound images and detect early indicators of congenital heart defects. Several computer vision models, including ResNet, DenseNet, and YOLO, are evaluated for their performance and limitations. Furthermore, an ensemble of these models is developed to improve accuracy and provide more reliable confidence estimates. Finally, a proactive medical support system is designed to integrate seamlessly into the clinical workflow, enabling physicians to benefit from automated analysis without disrupting their practice.

ABSTRACT. This paper introduces a convolutional neural network (CNN) model for the detection of synthetic speech, commonly referred to as deepfake audio. The model is trained on MFCC (Mel-Frequency Cepstral Coefficients) features extracted from the “for-2sec” subset of the “Fake-or-Real” (FoR) dataset, which includes both genuine and artificially generated voice recordings. By learning to identify subtle spectral and temporal irregularities in speech patterns, the CNN is able to effectively classify audio samples as real or fake. The proposed architecture consists of three convolutional blocks followed by a dense layer and a sigmoid output for binary classification. Experimental results demonstrate strong classification performance, with the model showing consistent precision and recall across both real and synthetic speech samples. Visual analyses using MFCC plots and Mel spectrograms further reveal interpretability on consistent acoustic differences between authentic and synthetic voices. The system offers a promising solution for addressing the growing threat of deepfake audio, particularly in applications that require fast and accurate verification of speech authenticity.

ABSTRACT. This paper presents a research study on objective evaluation of two state-of-the-art text-to-speech (TTS) systems, Glow-TTS and VITS, trained on the Romanian speech corpus MARA. Romanian, as a low-resource language, poses distinct challenges for neural TTS architectures. We assess both models using standard objective metrics: Mel-Cepstral Distortion (MCD), F₀ Root Mean Square Error (F₀ RMSE), and Short- Time Objective Intelligibility (STOI). Across a representative test set, Glow-TTS achieved average scores of 3.45 (MCD), 22.63 Hz (F₀ RMSE), and 0.87 (STOI), while VITS scored 3.47, 23.68 Hz, and 0.86 respectively. These objective values are competitive with state-of-the-art results on neural TTS and both systems exhibit high subjective naturalness, too. Glow-TTS shows slightly better prosodic accuracy, whereas VITS maintains strong spectral consistency. To ensure reproducibility and scalability, we employ only objective evaluation methods, which are particularly suited for under-resourced settings where subjective testing is less feasible

ABSTRACT. A crucial aspect of numerous computer vision applications, such as autonomous vehicles, mobile robots, and assistive systems is the ability to understand the environment. Humans are able to do this effortlessly perceiving and interpreting their surroundings through their visual system. Scene understanding remains a significant challenge in the field of artificial intelligence. Panoptic segmentation has emerged as a concept aiming to address this challenge, striving to enable artificial systems to understand scenes effectively through the analysis of visual information. This study presents a comparative analysis of two prominent panoptic segmentation models used in the automotive domain: EfficientPS, trained on the Cityscapes dataset, and Detectron2, trained on the COCO dataset. The comparison between EfficientPS and Detectron2 is based on their performance on Cityscapes dataset and on a custom dataset, SoVLite. Using SoVLite dataset allows to analyse the networks' performances on novel images. Finally, the study discusses the advantages and disadvantages of each neural network and explores potential future directions for development in the field of panoptic segmentation.

ABSTRACT. In-vivo monitoring of the heart activity is usually performed using electrocardiogram (ECG) signals. Accurate cardiovascular diagnostics require a clean, noise-free and artifact-free ECG. For this purpose, a wide range of noise and artifact detection techniques is being explored. This paper presents a feature-based AI/ML workflow for artifact detection, focusing on electrode contact noise and motion artifacts. The study was carried out on the publicly available MIT-BIH Noise Stress Test Database. The proposed feature-based ECG classification workflow employs the correlation coefficient matrix, generate from the mean, standard deviation, root mean square, entropy, and zero-crossing rate, to represent signal characteristics for artifact discrimination. Five supervised classifiers were evaluated for this purpose: Support Vector Machine, k-Nearest Neighbors, Decision Tree, Random Forest, and Artificial Neural Networks, based on accuracy, precision, sensitivity, specificity and F1-score. Random Forest demonstrated superior accuracy and generalization, with the best classification performance metrics accounting for 98.18% accuracy on the motion artifact dataset. The results demonstrate the applicability of lightweight feature-based ML models for robust ECG artifact detection.

ABSTRACT. Effective hospital resource management is critical for high-quality patient care. This study presents a predictive modeling approach within a multi-agent hospital management system to optimize patient scheduling, resource distribution, and waiting time estimation. A Random Forest Regressor predicts patient waiting times using demographic and medical urgency factors. A 30-day simulation showed a 20% reduction in average wait times, improved patient throughput, and optimized resource utilization. Experimental results confirm the efficacy of predictive modeling in streamlining hospital workflows. Graphical representations, downloadable as PNG images, illustrate patient urgency levels, waiting times, resource distribution, and doctor assignment efficiency.

ABSTRACT. Designing highly accurate lightweight semantic segmentation models is an essential task in the computer vision field due to the necessity to deploy these segmentation models in real-world scenarios. Developing optimized architectures is a challenging research direction since it needs to maintain a right balance between segmentation accuracy and inference speed. Therefore, we consider to rethink an advanced semantic segmentation model, called DeepLabV3+, into a more lightweight approach, enhancing its real-time performance. Through our method, we achieve better optimization across three important performance metrics, i.e., inference speed, GFLOPs (Giga Floating Point Operations), and the number of learnable parameters, while maintaining the accuracy at the same level. Our method is validated through extensive experiments conducted on two urban street scene datasets, namely Cityscapes and CamVid. In particular, on the Cityscapes dataset, our optimized model demonstrates a significant speed-up improvement from 276 FPS to 289 FPS on 512×1024 input resolution, a 75% reduction in learnable parameters, and a lower computational load of the model, 10.8 GFLOPs compared to 18.8 GFLOPs, and all these results occur while preserving the 74% accuracy level.

ABSTRACT. This study addresses automated analysis of crystal structures in microscopic saliva sample images to assess the salivary ferning phenomenon. The primary contribution is the design and implementation of a lightweight convolutional neural network, tailored to the complexity required for ferning recognition. This model features a simple architecture with a small number of parameters, optimized for computational efficiency and rapid training. Its performance was evaluated against a more complex and standard U-Net architecture, a widely accepted baseline in biomedical image segmentation. Both models were trained on a custom dataset of microscope images, which were annotated through a semi-automated process to segment three classes: crystals, non-crystalline areas, and background. The experiments, conducted using the Julia programming language and the Flux framework, demonstrated the effectiveness of the proposed approach. The results show that the lightweight model performed well. The findings suggest that for the specific task of recognizing the local, crystalline textures of saliva, a focused, computationally efficient architecture is more effective in contrast to a large and complex model designed to capture an extensive global context. The results of this study highlight the potential of customized deep learning solutions for specialized real-world applications.

ABSTRACT. This paper addresses the problem of detecting hazard-shaped markers on a support placed on glass frames for applications such as augmented reality and medical monitoring. The objective is to accurately locate these markers, which are placed in the upper portion of the face, within a rectangular region identified by a face detector. We describe two main methodological approaches: (1) a heatmap-based segmentation model that treats each marker as a channel to be localized, and (2) an object detection approach using versions of YOLO (e.g., YOLOv8, YOLOv11) at various scales. We employ both synthetic and real datasets, covering diverse capture distances. Extensive experiments are performed with architectures such as HRNet (modified to produce output channels at multiscale) and MobileNetV3, as well as YOLO variants. Our evaluation includes distance-based error metrics, indicating how precisely the center of each detected hazard marker aligns with the ground truth annotations. We provide both quantitative comparisons (through an aggregated table) and qualitative insight into the strengths, limitations, and potential future directions of each method.

ABSTRACT. Automatic car damage detection has been a topic of significant interest for the auto insurance industry as it promises faster, accurate, and cost-effective damage assessments. However, few works have gone beyond 2D image analysis to leverage 3D reconstruction methods, which have the potential to provide a more comprehensive and geometrically accurate representation of the damage. Moreover, recent methods employing 3D representations for novel view synthesis, particularly 3D Gaussian Splatting (3D-GS), have demonstrated the ability to generate accurate and coherent 3D reconstructions from a limited number of views. In this work we introduce an automatic car damage detection pipeline that performs 3D damage segmentation by up-lifting 2D masks. Additionally, we propose a simple yet effective learning-free approach for single-view 3D-GS segmentation. Specifically, Gaussians are projected onto the image plane using camera parameters obtained via Structure from Motion (SfM). They are then filtered through an algorithm that utilizes Z-buffering along with a normal distribution model of depth and opacities. Through experiments we found that this method is particularly effective for challenging scenarios like car damage detection, where target objects (e.g., scratches, small dents) may only be clearly visible in a single view, making multi-view consistency approaches impractical or impossible. The code is publicly available at: https://github.com/DragosChileban/CrashSplat.

ABSTRACT. Dynamic DNA nanotechnology enables applications in biosensing, molecular computing, and nanorobotics. Toehold-mediated strand displacement (TMSD) is a key mechanism in these systems. In this study, we replace a classical convolutional neural network (CNN) with a quantum convolutional neural network (QCNN), leveraging TensorFlow Quantum to predict TMSD rate constants (k_1). Using a dataset of 4450 DNA sequences with simulated k_1 values, we combine DNA-BERT embeddings with nucleotide availability and GC affinity features. The proposed QCNN model demonstrates strong performance, achieving an RMSE of 0.79 and R² of 0.74, comparable to the classical CNN. This demonstrates the promise of hybrid quantum-classical machine learning in the field of DNA nanotechnology.

ABSTRACT. Contrastive vision-language pretraining has shown strong performance in learning rich multimodal representations from image-text pairs. However, applying this paradigm to the medical domain presents significant challenges due to the complex visual patterns of medical conditions and the subtle, often ambiguous language used in radiology reports. To address these limitations, we propose CLIP-XRad, a weakly-supervised contrastive vision-language model specialized in chest X-ray interpretation. Our method introduces medical concept alignment into the contrastive learning process via a custom training objective that incorporates both instance-level and concept-level similarity, guiding the model to structure its embedding space around shared clinical semantics. Through extensive experiments, we demonstrate that CLIP-XRad achieves competitive performance in image-text retrieval, while also transferring effectively to downstream tasks such as multi-label classification and report generation. These results demonstrate that concept-aware contrastive pretraining can produce generalizable, semantically meaningful medical representations that provide a data-efficient solution for clinical applications.

ABSTRACT. Facial Expression Recognition (FER) is a complex task in machine learning, particularly in dynamic real-world settings. By enabling machines to accurately interpret human emotions, FER plays a pivotal role in advancing fields such as human-computer interaction, emotional state monitoring, and healthcare, especially for cases where human supervision is not practical or possible. This work presents a powerful yet adaptable real-time FER system anchored by a robust ensemble architecture. Our contribution unfolds in three phases: i) meticulous dataset curation and processing; ii) comprehensive model benchmarking and ensemble construction; iii) creation of a seamless real-time detection system that pairs a face detection module with a newly proposed ensemble model. By creating a composite dataset from multiple FER collections, we enhance system robustness and generalizability and create a base for future work in the field. We rigorously evaluated various image classification models to select optimal ensemble components, achieving state-of-the-art performance on the FER2013 benchmark and very good results on similar datasets. Our system demonstrates superior accuracy and generalizability, offering a versatile solution for real-world applications.

ABSTRACT. This paper investigates whether the integration of digital passports and gamification elements can improve engagement and learning outcomes in non-formal programming education, as implemented in the CodeClub UVT platform.

The platform extends the traditional CodeClub model by introducing a Digital Passport for progress visualization, a badge-based dual-layer gamification system, AI-generated personalized quizzes, and a digital portfolio to support project-based learning. These features were fully developed and tested within a standalone web application tailored for children aged 9 to 14.

Initial testing in real-world workshops demonstrated high engagement: over 90% of students appreciated the user interface, 87% completed quizzes, and more than half successfully used the portfolio feature. Results suggest that combining AI-driven personalization with gamified learning experiences can significantly enhance motivation and student participation in introductory programming education.