Abstract:
This paper investigates the benefits of further fine- tuning the HateBERT model for hate speech classification, addressing the challenges posed by class imbalance and linguistic subtleties inherent in hateful content. The fine-tuning process employs a gradual unfreezing schedule combined with a custom Focal Loss function to effectively handle imbalanced datasets and mitigate overfitting. Training is optimized using the AdamW optimizer with linear learning rate scheduling and gradient accumulation to maintain stable convergence. Evaluation on representative hate speech examples shows that the fine-tuned HateBERT achieves higher classification confidence compared to the original pre-trained model, improving the detection of subtle and implicit hateful expressions. Additionally, an LLM (Large Language Model)-based post-evaluation using Mistral-7B provides interpretable justifications and expert-style validation of the model’s predictions, enhancing transparency and trustworthiness. These results underscore the importance of task- specific adaptation and multi-model interpretability approaches to improve performance on complex and nuanced NLP tasks such as hate speech detection.

Abstract:
Online platforms are complex systems that influence the commercial, social, and political environment, debating important real-life topics, e.g., health, emigration, elections, climate change, etc. These online environments offer users freedom of expression through anonymous posting. In addition to their obvious advantages, some users abuse this freedom to spread harmful content, e.g., misinformation, propaganda, harmful conspiracy theories, or abusive, aggressive, and offensive speech. Automated detection techniques can effectively reduce the negative influence of antisocial behavior used by these malicious actors. In this article, we propose HARMLLAMA, a fine-tuned LLAMA2 model using LORA. The experimental results on two real-world datasets show that our model, HARMLLAMA, outperforms current state- of-the-art models in terms of Accuracy, Precision, Recall, and F1-Score.

Abstract:
Generative artificial intelligence models, including large language models and image generation models, are increasingly deployed in socially impactful domains. However, these models often exhibit social biases that can amplify stereotypes and produce harmful, discriminatory outputs. In this paper, we present a modality-comparative survey of social bias in text and image generation, structured around four components: benchmarks, bias identification, measurement, and mitigation. We systematically analyze methodological parallels and divergences across the two modalities, highlighting emerging research trends and identifying gaps. Finally, we map current image generation research efforts to the EU AI Act’s technical requirements, offering insights into how the community can advance towards more fair, safe, and trustworthy systems.

Abstract:
This paper presents an integrated medical platform that streamlines healthcare workflows by eliminating reliance on traditional document-based methods for handling patient records. Building on previous work on structuring unstructured medical data, the proposed extension now enables direct generation, editing, and storage of Electronic Health Records (EHRs) within a secure web interface. The platform supports dynamic form creation, customizable templates, role- based access, and cloud-based synchronization. To enhance usability, an intelligent assistant is integrated using a retrieval- augmented generation (RAG) pipeline. A locally trained NLP module handles intent classification and slot extraction, retrieving structured data from the medical database. Outputs are refined using OpenAI to deliver coherent, context-aware responses. The platform demonstrates improved efficiency, data consistency, and user interaction, aligning with digital health goals and EMRAM Stage 6+ criteria.

Abstract:
This paper introduces a novel approach for speaker anonymization based on cascaded speech-to-text (STT) and text-to-speech (TTS) Deep Neural Networks (DNNs), where the multi-speaker TTS component is driven by a Generative Adversarial Network (GAN) to anonymize the speakers’ identity and by a prosody transfer module to preserve the original prosody. The GAN creates artificial speaker embeddings that do not correspond to any real identity, enabling a high diversity of synthesized voices. The prosody preservation and speaker embedding generation for the TTS component are clearly distinct approaches: the embeddings are progressively generated from a low and compact latent space, while the prosody is directly cloned onto the synthesized speech. The experimental results show low cosine similarity between original and anonymized embeddings, with some minor degradation in linguistic fidelity due to the performance of the STT and TTS components. Additionally, the developed visual interface supports comparative analysis of audio features before and after anonymization, contributing to a qualitative assessment of system performance.

Abstract:
Deep learning models are widely used in the context of automated driving. Achieving the goal of performance in Automotive Embedded Targets presents numerous challenges due to the trade-off between accuracy and run-time. This paper addresses the task of finding the best candidate that fulfills System- on-Chip (SoCs) constraints using two-stage neural architecture search. We employed this approach to find a balance between accuracy and run-time. Our method can be applied in the context of optimizing deep learning models. Starting from a search space and a neural network, we can automatically create a super-model (by replacing all layers specified in the search space) that can be trained in the first stage of NAS. In the second stage, we use the already trained model to subsample architectures to find the best trade-off between accuracy KPIs and runtime KPIs. These two metrics are not arbitrary; they define the main constraints of SoCs and were used in the construction of the Pareto front from which the users can choose a configuration that satisfies their expectations. We conducted multiple experiments, but the most important ones are: optimizing a YOLOv8s for Texas Instruments (TI) SoC with 8 TOPS per Neural Network Accelerator and a multitask network for RENESAS SoC that has 3.4 TOPS. We ran YOLOv8s experiments on a public COCO dataset and managed to find a neural network that runs approximately 16% faster on TI, with a drop of less than 1.1% in mAP. For the Multitask Network, we conducted experiments on a custom dataset and found a neural network that runs 13% slower on RENESAS, but with an increase of 0.4% in overall evaluation metrics.

Abstract:
Accurate scene understanding from low-altitude aerial video cameras is a prerequisite for autonomous urban drones. However, current methods are confined to closed taxonomies and struggle with the high object density, scale variation, and viewpoint changes typical of UAV footage. We propose, to our knowledge, the first open-world video panoptic segmentation pipeline tailored for aerial perception, driven primarily by foundation models and enhanced with domain-specific pre- and post- processing filters. Built around the Grounded Segment Anything Model (Grounded SAM), our pipeline leverages the Decoupled Video Segmentation (DEVA) architecture. This enables scalable and memory-efficient temporal propagation, supporting robust segmentation and tracking over long video sequences with dense object distributions. A key innovation is an adaptive slicing mechanism inspired by the Slicing Aided Hyper Inference (SAHI) model, which adjusts the granularity of windowing based on spatial location. It applies finer slices to the upper part of the frame and coarser ones to the lower part, improving the detection of small aerial objects without compromising inference speed. Additionally, we introduce a class-aware slicing strategy that dynamically resizes SAHI windows depending on the predicted object class size, enabling effective segmentation of both small and large objects in the same scene while preserving open- vocabulary generalization. The use of Light HQ-SAM ensures efficient segmentation with minimal resource overhead, allowing the system to handle long videos with high object density. Operating fully autonomously, the pipeline extracts relevant textual labels directly from input frames using Llama 3.2 Vision, which seamlessly guides the segmentation and tracking components without requiring manual input.

Abstract:
This paper presents efforts to enhance driving safety through the analysis of on-road behavior and personalized driver feedback. It reviews recent literature focused on smartphone- based driving assistance applications, highlighting trends in sensor-driven behavior monitoring and in-app feedback. The hybrid mobile application DriverTracker, developed using the Ionic framework, is analyzed, detailing its architecture, implementation, and driver scoring algorithm. Experimental results suggest that DriverTracker can identify unsafe driving behaviors by integrating sensor data with contextual user input before and after the trip. While the current study is limited in scope, it offers a proof of concept for combining subjective and objective metrics in driver behavior analysis.

Abstract:
This paper proposes a novel processing pipeline for real-time gaze tracking utilizing standard webcams in unconstrained environments. First, the study evaluates three gaze estimation estimators: the GazeTracking library, and two L2CS-Net deep learning models that have been trained on the Gaze360 and MPIIGaze datasets. Then, several improvements are proposed based on a user-specific calibration on a 5x4 grid, after which gaze data is mapped to screen coordinates using two regression approaches: linear regression is based on gaze data alone, whilst Ridge regression is dependent upon both gaze data and head pose data. A post-processing pipeline incorporating a moving median and a Kalman filter has been demonstrated to stabilize the predicted gaze point. This research demonstrates an effective, affordable pipeline for interactive systems and provides a comprehensive analysis of the trade-offs between the evaluated technical solutions.

Abstract:
Fine-grained change detection in structured remote sensing environments, such as agricultural fields, requires models that can reason over semantically coherent regions while preserving spatial precision. Existing methods often rely on convolutional or transformer-based architectures with uniform tokenization, limiting their ability to capture changes aligned with real- world semantic structures. In this paper, we propose SFCNet, a region-aware change detection framework that leverages explicit field-level segmentation masks to guide both token aggregation and spatial decoding. Our method introduces a Field-Aware Region Encoder (FARE) that aggregates features within each semantic region, producing interpretable region tokens. These are processed through a slot attention mechanism to model temporal differences across structured units. A re-projection decoder then restores field-wise predictions to pixel-space using the original geometry. We evaluate SFCNet on the CLCD dataset, where it outperforms prior state-of-the-art models by +3.31%.

Abstract:
This study presents a machine learning-based approach to optimize the design and thermal performance of Electronic Control Units (ECUs) by predicting the Heat Transfer Coefficient (HTC), a key metric in thermal management. ECUs play a critical role in vehicle systems, and efficient thermal control is essential for their reliability. Traditional CFD simulations, while accurate, are time-consuming and resource- intensive. To address this, the study uses data from Computational Fluid Mechanics (CFD) simulations and applies Machine Learning (ML) models to predict HTC based on geometric and thermal inputs. Results show that these models can streamline the design process, supporting rather than replacing simulation engineers, and enabling faster evaluation of viable configurations. This approach promotes more agile, efficient, and sustainable ECU design in the automotive industry.

Abstract:
This paper describes an improved solution for the problem of accurate detection and classification of hand-drawn logic circuit components. The solution builds upon previous work which has shown promising initial results in logic gate classification and experimented with connection detection. Herein, we describe an improved connection detection solution, which is based on a YOLO classifier, and evaluate its performance. Additionally, we expand upon the initial collection of classes, adding more complex logic circuit components. The result is an in-memory model that can be used to test the circuit and verify its correctness. Throughout the paper, we provide an overview of our solution and describe a series of test scenarios we devised to evaluate the proposed solution.

Abstract:
Machine learning (ML) has become an increasingly important component in integrated circuit (IC) development, particularly in pre-silicon verification. Traditional methods often face challenges such as slow processing, high resource demands, and difficulties in achieving comprehensive design space coverage. In contrast, adaptive ML techniques concentrate on critical regions of the design space while simultaneously ensuring extensive coverage, enabling faster convergence and lower simulation costs. By employing stopping criteria based on the convergence of a cost function, these methods enhance both accuracy and computational efficiency. As IC complexity continues to rise, the integration of ML-based approaches into the manufacturing flow offers a practical and scalable path toward robust, cost-effective development workflows.

Abstract:
This research presents a novel clustering- enhanced methodology for industrial anomaly detection that demonstrates significant performance improvements over conventional approaches applied directly to raw sensor data. The investigation employs Self-Organizing Map clustering to systematically partition multivariate operational data from PET bottle manufacturing processes, creating distinct contextual environments that enable more precise anomaly identification. The methodological framework centers on the strategic integration of unsupervised clustering as a preprocessing stage for anomaly detection algorithms. By clustering complex sensor data containing power consumption, we establish operational contexts that facilitate accurate detection of anomalous behaviors within specific manufacturing states. This approach addresses limitations in traditional anomaly detection methods when confronted with industrial sensor data. Our experimental design incorporates multiple advanced anomaly detection algorithms like Density-Based Spatial Clustering of Applications with Noise, Isolation Forest, Copula-Based Outlier Detection, Empirical Cumulative Distribution-based Outlier Detection and Learnable Unified Network for Anomaly Recognition applied comparatively to both raw sensor data and clustered operational contexts. The systematic analysis reveals substantial performance increases when anomaly detection operates within clustered environments rather than on unprocessed data streams.

Abstract:
In modern distributed systems, the Remote Procedure Call (RPC) protocol is essential for inter-machine communication. Nevertheless, it also provides attackers with a stealthy pathway for lateral movement (pivoting to other machines after an initial compromise). Existing lateral movement detection efforts often focus on authentication logs or Remote Desktop Protocol sessions, potentially missing stealthier threats that leverage RPC. We develop an unsupervised anomaly detection framework using the Isolation Forest algorithm, which learns standard RPC traffic patterns on a per-host basis and flags deviations indicative of attacks. To evaluate our method, we constructed a simulation framework that generates realistic benign RPC traffic, and we implemented lateral movement scenarios using real- world attack tools. Experimental results show that our approach effectively detects anomalous activities, highlighting its potential for proactive RPC-based threat identification.

Abstract:
This paper surveys recent mechanisms for inducing compositional behavior in neural networks, with a focus on their potential to support explainability. We review approaches based on architectural design, including external stacks, modular networks, and relational attention, as well as training strategies such as meta-learning, curriculum design, and iterated learning. These methods promote the formation of internal representations that support systematic generalization. We also analyze how such representations reduce feature entanglement and enable component-based explanations. Thus, compositionality appears to be not only a functional objective but a desired design principle for explainable neural systems.

Abstract:
Generating optical pulses by semiconductor lasers with duration below 10 picoseconds is a significant area of research in photonics, enabling advancements in high-speed optical communications and ultrafast spectroscopy. In this paper, we dealt with the problem of short pulses generation by gain-switched DFB lasers. We addressed the problem in two separate ways. The first one was based on running simulations for various values of the following features: injected current, optical confinement factor per quantum well, number of quantum wells and coupling coefficient. As a result, we built a synthetic dataset of roughly 14,995 examples, which can be used for further research. We also explored several Machine Learning models aimed at building a function for predicting peak power, full width at half maximum, total energy and pulse energy respectively. We explored Linear Regression, XGBoost, Random Forest and Tabular Prior-data Fitted Network. 4-fold cross validation was applied, with R2 Score used as a performance metric. For each target, TabPFN2 lead to the best model, with the following performances: 96.97% (Peak power); 88.33% (full width at half maximum); 99.88% (total energy); 97.50 (pulse energy).

Abstract:
Spike sorting is the process of assigning extracellularly recorded action-potential waveforms to their putative neuronal sources; it is an inherently unsupervised yet mission-critical step in large-scale electrophysiology. Handcrafted features and clustering heuristics have long limited spike-sorting accuracy, portability and scalability. In this work, we reformulate spike sorting as a self-supervised representation-learning problem and introduce a reconfigured Self-Supervised Audio Spectrogram Transformer (SSAST). The proposed multi-branch architecture converts one-dimensional neural signals into two-dimensional representations through three complementary pathways—Continuous Wavelet Transform (CWT), short-time Fourier transform (STFT) and direct signal remodelling—and fuses mid- and high-level features by means of MultiLayer Pooling Fusion. Relative to the original SSAST, the modified model expands the embedding dimensionality from 768 to 3072, boosts classification accuracy from 64.4% to 76.3%, and shortens training by 24%. These results demonstrate that multi-scale self-supervised learning can generate reliable labels and substantially improve spike- sorting performance.

Abstract:
Mutations in the TP53 gene are among the most prevalent alterations in human cancers and are known to influence therapeutic responses. Accurate prediction of drug sensitivity based on TP53 mutation profiles remains a major challenge in precision oncology. In this study, we propose a quantum-enhanced framework for genomic classification using Quantum Boltzmann Machines (QBMs) applied to quantum- encoded DNA sequences. Exon 5–8 fragments of TP53 containing both wild-type and clinically relevant mutations were encoded using angle-based quantum rotations and input into a QBM architecture trained to predict binary drug response (sensitive vs resistant) based on half-maximal inhibitory concentration (IC50) thresholds. Simulated results demonstrate that the QBM achieved over 91% accuracy in classifying drug sensitivity, outperforming classical Boltzmann models and revealing interpretable latent features. This work showcases the potential of QBMs in modeling sequence-driven phenotypes and supports their application in biomarker discovery, drug response prediction, and quantum-enabled precision medicine.

Abstract:
String processing is a fundamental component of numerous applications in computer science, with particularly critical roles in areas such as bioinformatics, text mining, and data compression. In bioinformatics, efficient string matching underpins essential tasks like genome alignment, motif search- ing, and sequence assembly. However, traditional CPU-based implementations of string algorithms often struggle with the scale and complexity of modern data sets. This review explores the acceleration of classical string algorithms using NVIDIA’s CUDA platform for GPU computing. We examine and compare the performance and parallelization strategies of Knuth-Morris- Pratt, Boyer-Moore and Burrows-Wheeler Transform when im- plemented on GPUs. The review analyzes how each algorithm’s structure affects its adaptability to CUDA’s parallel architecture, and evaluates their speedup, scalability, and memory efficiency across varying input sizes.

Abstract:
This study presents an algorithm designed for the Controller Area Network (CAN) protocol, which is capable of receiving both standard CAN frames and CAN Flexible Data- Rate (CAN FD) frames to identify active Diagnostic Trouble Codes (DTCs) within a vehicle. Given the pivotal role of the automotive sector in today’s industrial landscape, the research emphasizes autonomous diagnostics, predictive maintenance, and prescriptive maintenance methodologies — approaches that are critical not only for modern vehicles but also for broader Industry 4.0 contexts. Experimental results demonstrate a minimum im- provement of 15% in data validation and interpretation accuracy. This enhancement is particularly significant in industrial and real-world applications, where the timeliness and reliability of malfunction detection are essential for operational efficiency and safety.

Abstract:
In this work, we propose an innovative knowledge distillation framework tailored for promptable transformer-based segmentation models, focusing on Segment Anything Model 2 (SAM 2). Powerful models, such as transformer-based ones, which require large-scale datasets and extensive computation for good performance, are difficult to train, even with limited data, often yielding uncertain segmentation results across different use cases. We address this challenge by proposing a distillation strategy that allows a student model to inherit structured knowledge from a strong pre-trained teacher with the same architecture, even when trained on small datasets. The method combines hard supervision from ground-truth masks with soft guidance from the teacher’s logits, helping the student learn meaningful structures early in the training process. To address the challenge of noisy or overconfident predictions from the teacher, we introduce a confidence-aware agreement mask that modulates the soft loss and provides guidance for the student to focus on regions where the teacher is more confident. The work is extended with token-level distillation to enhance structural fidelity in student representations. Experiments demonstrate the effectiveness of our method in low-data regimes, with the student model achieving a notable reduction in the total loss (from 15.17 to 8.84) and up to a 20% improvement in Intersection- over-Union (IoU) across various inference scenarios. Qualitative results further support these findings, showing visibly cleaner segmentation outputs with reduced noise, indicating improved generalization under data-scarce conditions.

Abstract:
This paper proposes a hybrid vehicle detection method for LiDAR point clouds, combining geometric algorithms for segmentation and bounding box fitting with a lightweight machine learning classifier. Our approach is based on an L-shape fitting method but also integrates spatial filtering and adaptive clustering to enhance robustness in complex urban scenes. The pipeline consists of ground removal, density-based clustering, vehicle cluster identification, and refined L-shape fitting, which efficiently approximates vehicle footprints. A Random Forest classifier enhances robustness by filtering non-vehicle clusters. Experiments conducted on the SemanticKITTI dataset show that the proposed method achieves a mean precision of 68.16%, while maintaining a significantly lower computational cost. The obtained results are comparable to deep learning methods, while preserving real-time performance and interpretability, making the approach suitable for lightweight autonomous driving platforms.

Abstract:
Perception in aerial environments depends on object detection and semantic segmentation to achieve scene understanding. However, the limited availability of semantically annotated datasets poses a challenge due to the extensive manual effort needed for their creation. This work proposes a customized pipeline designed to generate high-quality semantic masks from object detection bounding boxes by integrating the Segment Anything Model (SAM), a suite of image processing techniques, and adaptive prompting methodologies. The input prompts represent either foreground objects or background regions and are refined through additional image processing operations to improve quality and mask accuracy. The primary contribution of this work is the use of point-based prompts to accurately guide the class-agnostic SAM model in delineating precise semantic boundaries. Furthermore, a genetic algorithm is employed to intelligently select a reliable and informative set of object points, thereby enhancing the robustness and accuracy of the segmentation process. The proposed method was evaluated using a manually annotated subset of the AU-AIR dataset, focused on vehicle instances, achieving a segmentation score of 84.62% IoU. This represents an 18% improvement over the baseline UNetFormer model trained on the UAVid dataset.

Abstract:
This paper presents an innovative and interpretable LiDAR-based architecture for weather condition detection and identification, leveraging the Gray-Level Co-occurrence Matrix (GLCM) and a dedicated AI model. By exploiting the unique spatial characteristics of LiDAR point cloud data, the proposed approach enhances the accuracy and robustness of detecting three distinct weather conditions: clear, rainy, and foggy. The LiDAR point cloud is preprocessed and projected into a discretized map, from which GLCMs are computed and used as input for model training. In addition, interpretability is a core feature of the method, addressed through two complementary techniques. First, an ante-hoc analysis estimates the 2D conditional probability distribution of GLCM features given the weather class prior to training. Second, a post-hoc visualization is enabled via a reshaped weight map that reveals the information learned by the model. The model is evaluated on the RADIATE dataset and benchmarked against three state-of-the-art weather classifiers. Results demonstrate the effectiveness of the LiDAR–GLCM approach and its strong potential for real-time, embedded weather recognition systems.

Abstract:
Natural disaster assessment relies on accurate and rapid access to information, with social media emerging as a valuable real-time source. However, existing datasets suffer from class imbalance and limited samples, making effective model development a challenging task. This paper explores augmentation techniques to address these issues on the CrisisMMD multimodal dataset. For visual data, we apply diffusion-based methods, namely Real Guidance and DiffuseMix. For text data, we explore back-translation, paraphrasing with transformers, and image caption-based augmentation. We evaluated these across unimodal, multimodal, and multi-view learning setups. Results show that selected augmentations improve classification performance, particularly for underrepresented classes, while multi-view learning introduces potential but requires further re- finement. This study highlights effective augmentation strategies for building more robust disaster assessment systems.

Abstract:
A robust and efficient system for face identification and recognition in video streams and image sequences is presented in this paper. Leveraging state-of-the-art face detection and margin-based embedding models, the proposed framework accurately extracts discriminative facial features and performs scalable similarity search within large identity databases. A key component of the system is a semi-automatic face registration module that combines automated detection with user-guided selection to incrementally build high-quality face datasets. Designed to operate in real-time, the system achieves a balance between high recognition accuracy and low latency, while maintaining robustness against common environmental variations such as changes in lighting, pose, and occlusion.

Abstract:
Guaranteeing privacy in facial recognition remains an open challenge due to the sensitive nature of biometric data. This paper presents the development of a biometric authentication system based on facial recognition, designed to ensure complete data confidentiality by integrating fully homomorphic encryption. Using this technique, sensitive data remains encrypted throughout the entire processing pipeline and is never exposed in plaintext on the server. The system is based on a Siamese neural network trained to compare two facial images and determine whether they belong to the same person, with all computations performed directly on encrypted data using the CKKS scheme, which enables approximate arithmetic over encrypted real numbers. The performance of the encrypted system is evaluated against the unencrypted baseline model, highlighting the trade-offs between security and computational efficiency.

Abstract:
This paper investigates the detection of black rot disease in grapevines using both flash and natural outdoor imagery, emphasising real-world field conditions over controlled laboratory environments. Leveraging a dataset captured from an Unmanned Ground Vehicle (UGV), the study evaluates the robustness of deep learning models under few-shot learning scenarios using flash and natural light. A hybrid architecture combining a Feature Pyramid Network is employed, with models initially trained on annotated source data and adapted to unlabeled target domains through transfer learning. The results demonstrate that few-shot learning performance in visual disease detection is significantly enhanced when adaptation is guided by a prior model trained on broader symptomatology, underscoring the importance of generalised representations in few-shot conditions.

Abstract:
Multi-label classification methods have gained im- portance in the context of modern machine learning applications, as they can handle the complexity of a data point concurrently being associated with multiple semantic concepts; however, standard approaches often face the challenges of data scarcity and imbalance. These two issues are driven by the rapid advances in real data collection, where repositories can present large volumes of information for specific label combinations, while under- representing others. In this research paper, we present a multi- label classification architecture: ContrastRelevance. Composed of a contrastive learning model and a static binary relevance structure, it is designed for problems affected by scant data.

Abstract:
Automated multi-label electrocardiogram (ECG) diagnosis promises accessible cardiovascular care, but clinical adoption is challenged by deep learning models' size, interpretability, and poor performance on rare conditions. This paper presents an end-to-end framework addressing these issues through two novel, compact convolutional neural network (CNN) architectures—an Efficient CNN (~20k parameters) and a Compact MiniResNet (~35k parameters)—trained on the 21,837-record PTB-XL dataset. Our methodology integrates (1) a robust class weight strategy for label imbalance, (2) waveform- aligned Grad-CAM for clinically relevant visual explanations, and (3) post-training, per-class decision threshold optimization to significantly improve sensitivity on minority classes. Models are validated for edge deployment via full INT8 quantization. Our compact models achieve diagnostic performance comparable to multi-million parameter ResNet baselines (MiniResNet Macro-AUC of 0.961, Macro F1-score of 0.70). Threshold optimization boosted recall on the rarest class (Hypertrophy) from 0.17 to 0.52. Post-quantization, model size is reduced by 87% (to 65 KB) with negligible performance drop, and inference latency by 18x on a CPU, confirming edge suitability. This work provides a reproducible blueprint for developing trustworthy, lightweight, and interpretable AI- powered cardiology assistants.

Abstract:
Automatic detection of epileptic seizures based on EEG signals represents a major challenge in neuroscience and AI- assisted medical applications. Currently, over 50 million people worldwide live with epilepsy, and a significant proportion of them experience recurrent seizures that can occur unpredictably and severely affect their quality of life. For this reason, we propose EnfXAI, a hybrid system that integrates a deep learning model (CNN), an explainability component (SHAP), and a fuzzy logic module for risk estimation. The CNN model is trained on an EEG dataset segmented into one-second windows, achieving an accuracy of over 98% in classifying seizure and non-seizure windows. SHAP is then used to identify the EEG features with the highest positive contribution to seizure prediction. These features are subsequently processed by a fuzzy logic system that provides an interpretable seizure risk estimate-classified as low, medium or high. In this way, we contribute to the early and interpretable identification of epileptic seizures, helping affected individuals benefit from faster and more personalized interventions. As a result of the analysis, the test data was split into two categories: at-risk (693 individuals) and not at-risk (494 low-risk and 1113 medium-risk individuals). The final output is both a highly accu- rate and easy-to-understand prediction, with strong potential for application in real-world clinical contexts. The proposed system represents a significant step toward integrating explainable AI (XAI) into neurological diagnostics.

Abstract:
Methods for neuronal spike feature extraction and clustering are a fundamental part of analyzing neuronal activity and understanding how the brain processes information. Extracting relevant features from neuronal signals, time-frequency transforms such as the Wavelet Transform or the Superlet Transform, provides a detailed description of the shapes and properties of spikes. This step can help distinguish neural signals from noise and preserve relevant information about neural activity. The clustering of spikes by algorithms such as K-Means or probabilistic models such as the Space Breakdown Method facilitates their grouping according to their neuronal source. This process can further our understanding of neural interactions and help us in analyzing the functioning of brain networks. Our work aims to contribute to the improvement and comparison of neural data processing methods, with a direct impact on neuroscience research.

Abstract:
Wearable sensors (e.g., accelerometers, gyroscopes) combined with machine learning (ML) methods offer a promising solution for remote monitoring of physical rehabilitation exercises. However, challenges arise given the high variability in how patients perform the same exercises, combined with the scarcity of labeled data. This limits the effectiveness of standard supervised learning approaches. This paper analyzes and quantifies key aspects of developing generalizable physical exercise recognition models through the lense of ML training procedures. We make four contributions: (i) evaluating time-domain data augmentations on model performance, (ii) quantifying inter- person variability using within-dataset incremental fine-tuning, (iii) assessing transferability across datasets with encoders pre-trained in a supervised manner, and (iv) comparing supervised and unsupervised pretraining strategies using free-living and exercise-specific motion data.

Abstract:
This paper presents an automated framework for anatomical landmark detection and multi-class segmentation of the femur from CT images, designed to support preoperative planning in total hip arthroplasty (THA). Using a manually annotated subset of the TotalSegmentator dataset, we trained and evaluated two deep learning architectures—Attention U-Net and V-Net 2D—on 11 distinct femoral substructures, including head, neck, and trochanters. The proposed pipeline integrates segmentation with automatic extraction of key anatomical landmarks and enables precise computation of the femoral neck resection plane directly from model predictions. Both networks achieved high Dice scores (above 0.96), with Attention U-Net offering faster convergence and V-Net showing greater stability on imbalanced classes. Our results demonstrate that deep learning-based segmentation combined with landmark detection can significantly enhance accuracy and reproducibility in THA planning workflows.

Abstract:
The diagnosis of musculoskeletal disorders such as scoliosis is highly dependent on a correct measurement of the Cobb angle. We present a novel adaptation of the I-MedSAM model for approaching this task, by tailoring it for instance segmentation of vertebrae in spinal X-rays. Our approach introduces a detection head for generating bounding box prompts for I-MedSAM’s mask decoder and modifies the original loss function to better capture small anatomical structures such as vertebrae. On a limited clinically annotated dataset, our method achieved a mean absolute error (MAE) of 3.382, and a symmetric mean absolute percentage error (SMAPE) of 11.863, approaching the precision of manual measurements. These results are attained through training the segmentation model with regard to computational limitations and highlight the potential of our model as a low-cost AI-assisted diagnostic tool in clinical settings. Future work will focus on improving segmentation quality and dataset diversity through pseudo-labeling.

Abstract:
Colonoscopy is a widely used method for mucosal inspection and lesion detection. Accurate measurement of withdrawal time during colonoscopy is critical to ensure adequate mucosa inspection and improve the rate of adenoma detection. In this paper, we present a two-stage pipeline that combines a frame filter, based on a voting system (frame sharpness, structural richness, and color context), with a deep learning model for anatomical colon region classification to accurately estimate the withdrawal time for each colon region. Specifically, we introduce an adaptive thresholding mechanism to robustly identify non- informative frames, providing reliable frame exclusion, enhancing subsequent analysis. Qualitative results indicate that our model effectively tracks the colonoscope’s position through the colon regions during the withdrawal phase with reasonable temporal alignment, while managing uncertain frames.

Abstract:
A robust and efficient system for face identification and recognition in video streams and image sequences is presented in this paper. Leveraging state-of-the-art face detection and margin-based embedding models, the proposed framework accurately extracts discriminative facial features and performs scalable similarity search within large identity databases. A key component of the system is a semi-automatic face registration module that combines automated detection with user-guided selection to incrementally build high-quality face datasets. Designed to operate in real-time, the system achieves a balance between high recognition accuracy and low latency, while maintaining robustness against common environmental variations such as changes in lighting, pose, and occlusion.

Abstract:
Providing a promising solution for the detection of diabetic retinopathy (DR), this study introduces Inception Ret- Net, a new CNN architecture designed to support early diagnosis and prevent vision loss in diabetic patients. Built on InceptionV3, it incorporates customized attention modules (HAM and CBAM), enhancing feature extraction from retinal fundus images that are preprocessed with CLAHE and contour cropping. Evaluated on the EyePACS public database and combined with a custom dataset, the model surpasses baseline architectures such as state-of-the-art models. Trained with a Weighted Loss Function and Nadam optimizer, it effectively addresses class imbalance. Evaluation metrics include Accuracy, Precision, Recall, F1-Score, and AUC. This hybrid approach provides a robust framework for automated DR screening, with potential for clinical adoption to improve patient outcomes.

Abstract:
Accurate depth estimation from cameras is a highly important task for autonomous systems. Existing monocular depth estimation methods span a wide range of approaches, from supervised techniques which leverage labelled datasets, to self-supervised methods which utilize video sequences without explicit depth annotations. Additionally, stereo vision solutions provide absolute depth measurements using the disparity maps obtained via stereo correspondence. This paper proposes a depth fusion system combining Monodepth2, MiDaS and stereo vision using semantic-aware scaling and error-aware selection. The original contributions of our work are three-fold. First, our approach proposes an original method of scaling the relative disparities from monocular depth estimation using semantic segmentation and an original depth discretization technique. Then, an original method is proposed for combining self-supervised and supervised approaches using the probability of information obtained from learning the error each type of monocular depth estimation method produces. The final contribution consists in the presentation of the fault tolerant system used to reconstruct the scene. The proposed approach has been tested on the KITTI dataset and highlights the effectiveness of combining these complementary methods showing good results even in situations where individual methods would fail.

Abstract:
Space surveillance using optical instruments involves detecting faint streaks from long exposure images. As the space object moves, it creates a line segment trajectory in the image space, against the fixed star background. Clouds, light pollution, or the weak illumination of the object by the Sun may lead to a very small signal to noise ratio. The YOLO family of neural networks is a lightweight solution for detecting various objects from images. Starting from existing streak detection solutions, we have iteratively generated training datasets by a combination of manual and automatic annotation and augmentation. The bounding box output of the neural network detector must be further processed to detect the satellite streak’s line segment. The unfiltered output of the neural network, containing multiple overlapping boxes, can be used to indicate the position of the line, without any additional processing cost. The complete processing pipeline can detect faint streaks from any type of input image, with real time performance.

Abstract:
This paper presents a comprehensive machine learning framework for crop yield prediction that integrates satellite imagery, meteorological data, and ground-truth yield labels. We introduce two novel MMST-ViT-based model variants—AgriFormer-Lite and AgriFormer-RoSim—each designed for different levels of agro-climatic complexity. In addition, we develop a custom CropNet-RoSim dataset that simulates Romanian agricultural conditions. In experiments across four major crops and two data scenarios, both variants outperformed the MMST-ViT baseline. AgriFormer-Lite excelled in low-diversity settings, while AgriFormer-RoSim achieved the best results in high-variance conditions.

Abstract:
Camouflaged Object Detection (COD) involves identifying and segmenting objects that become part of the background. It is a complicated task for Computer Vision, which requires techniques able to exploit visual features beyond human eye capabilities. Traditional methods using pixel-wise local analysis in the RGB color space are usually insufficient for this task. After comprehensive testing, we noticed that Mask R-CNN models, provided with a supporting attention map using Discrete Cosine Transform (DCT) information, proved to be more accurate in detecting camouflaged objects. We developed five original models based on Mask R-CNN with different backbones, including ResNet-50, Res2Net, ResNeXt. Additionally, the models were enhanced by inserting a module based on DCT, which generates attention maps using the Self-Attention mechanism from Transformers. We used pre-trained backbones and trained our models on COD10K dataset. The models were evaluated using standard COCO IoU-based metrics. We also adopted Structure Measure to assess a superior qualitative examination specific to the segmentation task. Experiments conducted on the COD10K dataset demonstrate that using a ResNet-50 backbone combined with our DCT Self- Attention module improves the performance of standard Mask R-CNN with approximately 0.08 on AP@0.5 and 0.03 on Structure Measure.

Abstract:
This paper presents the design and implementation of an automated traffic monitoring and violation reporting system integrating advanced computer vision and machine learning technologies. The system unifies vehicle detection based on convolutional neural networks, optical character recognition for license plate reading, and real-time speed estimation into a scalable, efficient, and autonomous platform. The architecture incorporates multiple image preprocessing steps to enhance recognition accuracy and employs a multi-threaded design for efficient parallel processing of video streams. A key innovation is the homography estimation of camera parameters to adapt to camera motion, supported by a top-down projection method for precise distance and speed estimation. Reliable communication and automated violation reporting are achieved through a distributed messaging protocol that ensures scalable data transfer and automatic alert delivery to authorities. The comprehensive integration of these components into a single cohesive system reduces complexity and enhances maintainability, providing an effective solution for intelligent traffic surveillance and enforcement within modern transportation infrastructure.

Abstract:
Low-resource languages face significant challenges in the development of robust natural language processing (NLP) tools due to the scarcity of standardized corpora and linguistic resources. Tokenization is a critical step in natural language processing, yet existing tokenization algorithms often struggle with agglutinative and low-resourced languages due to their rich morphological structure and limited training data. In this paper, we present the creation of a corpus for Crimean Tatar, a severely under-resourced Turkic language. To address orthographic inconsistencies and dialectal spelling variations, we utilize bilingual dictionaries to normalize and consolidate different variants of the same word. Leveraging the linguistic similarity between Crimean Tatar and Turkish, we utilize a large Turkish corpus alongside a smaller Crimean Tatar corpus to adapt and improve subword tokenization algorithms such as Byte-Pair Encoding (BPE), WordPiece, and Unigram. We present a fine-tuning algorithm that incorporates a golden set of annotated tokenizations and a curated Crimean Tatar vocabulary to adjust token boundaries. To further address context-sensitive morpheme boundaries, where some morphemes should be merged or separated depending on usage, we model tokenization as a character-level sequence labeling task. A BiLSTM model trained on the golden dataset is introduced, assigning ”B” (begin) and ”I” (inside) labels to characters to predict token splits. This approach demonstrates improved handling of context-dependent segmentation and outperforms traditional tokenization methods. We provide detailed analyses of the corpus composition, normalization strategies and tokenizers creation and evaluation. Our results demonstrate that a larger corpus of a very similar agglutinative language, in this case Turkish, can improve the performance of the tokenizer, but more important is the usage of a vocabulary. However, a model which takes into account the context of characters can better tokenize words composed of morphemes than tokenizers based on learning merging rules.

Abstract:
In this paper, we propose a solution that leverages knowledge graphs to model and compare cooking recipes, aiming to bridge the gap between free-form user instructions and a structured recipe database. By integrating a hierarchical, domain- specific knowledge graph with inductive graph convolutional networks (for scalable embedding) and a fine-tuned multilingual transformer (for precise information extraction), our approach excels in both efficiency and matching accuracy. We achieve 97.91% accuracy on embedded recipes within the semantic network, yielding a maximum similarity of 0.54 for different versions of the same dish, and reach up to 0.32 similarity for dishes not present in the database. The motivation for this work lies in the need for end users, especially in low-resource or multilingual settings, to retrieve reliable, relevant recipes without manual browsing or query formulation. Our system addresses this by parsing user-provided recipe text in six languages, embedding both user and candidate recipes into a shared vector space, and computing similarity results to return the best match.

Abstract:
his research examines and compares the effectiveness of Long Short-Term Memory (LSTM) models and Large Language Models (LLMs) for stock trend forecasting by integrating sentiment analysis of financial news and Google Trends data. Advanced machine learning techniques are required to process large datasets and uncover complex patterns, and recent developments in LLMs present new opportunities for financial prediction. An LSTM neural network was trained on historical stock data, sentiment scores, and Google Trends data, with its performance compared against four state-of-the-art LLMs: Phi- 4, Llama-4 Maverick, Gemini 2.5 Flash, and DeepSeek R1, as well as a Vector AutoRegressive model. The study evaluates performance across four major stock tickers using comprehensive metrics. The LSTM model showed significantly lower error metrics compared to the VAR baseline, while LLMs demonstrated superior performance when provided with detailed prompts, achieving 0.32% sMAPE with perfect directional accuracy. The study used hyperparameter tuning, early stopping, and model checkpointing to enhance LSTM performance, while prompt engineering proved critical for LLM success. The findings suggest that both classical LSTM approaches and modern LLMs can effectively improve stock market predictions, with LLMs showing particular strength in processing multi-modal financial data.

Abstract:
Adverse event detection for non-pharmaceutical substances, such as dietary supplements, herbal remedies, and alternative therapies, remains a major blind spot in current pharmacovigilance systems, which are primarily designed for prescription drugs. This research addresses that gap by introducing a high-precision matching framework that links scientific literature from PubMed with real-world safety reports from the FDA Adverse Event Reporting System (FAERS). The proposed approach incorporates a robust pre-processing pipeline, followed by a three- layer matching engine based on rule-based substance overlap, metadata alignment, and temporal validation. The result is a curated set of 560 article–report pairs obtained by applying our multi-stage matching pipeline to two datasets: a filtered subset of 38,978 PubMed articles that mention non-pharmaceutical substances, drawn from the 2024 PubMed baseline distribution and 135,848 adverse event reports from the FAERS for Q3 2024 (July– September). This represents the first structured dataset connecting peer-reviewed research with reported adverse events for non- pharmaceutical compounds. Experimental results show that while raw rule-based matching produced 670 links, our enhanced pipeline reduced noise and improved semantic accuracy, prioritizing reliability over quantity. This dataset lays the groundwork for safer, data-driven integration of alternative medicine into modern pharmacovigilance efforts. Moreover, results so far serve as an important first step toward making available a benchmark dataset for evaluating non-pharmaceutical adverse event extraction and article-report linking systems. This work provides a reproducible, high-precision framework for linking literature and adverse event reports, addressing challenges not covered in prior studies.

Abstract:
When implementing an automated support system, organizations tend to rely on complex, computationally intensive architectures. This paper presents the results we obtained when exploring the possibility of developing a cost-effective question answering framework for this purpose. We explored how various models, context configurations, and personas (instructions that guide the system in terms of answer purpose and/or tone) impact customer satisfaction regarding the respective answers. Three Llama-3, 7 billion parameter version variants, fine-tuned across different combinations of training parameters, extra context provision and personas were evaluated. The solution obtained was implemented as a chatbot focused on questions regarding a specific university. Its performance was assessed based on the customer satisfaction scores obtained from university students. Results demonstrate that fine-tuned models with 1000 iterations (the point at which the model starts to overfit slightly) achieved up to 85% customer satisfaction when combined with both extra context and persona features, compared to 25% for the base model alone. These findings suggest that effective specialized chatbots could be implemented without complex Retrieval Augmented Generation architectures and provide practical guidance for organizations which aim for simplicity, cost-effectiveness and replicability in their systems’ development.

Abstract:
The increasing digitalization and networking of safety-critical infrastructures requires robust software management and secure update mechanisms. This case study investigates the transferability of established software update and cybersecurity standards from the automotive industry— namely ISO 24089, ISO/SAE 21434, and ISO/IEC 27001—to the domain of fire alarm systems. While vehicles and fire alarm systems differ fundamentally in terms of mobility and user interaction, they share comparable risk profiles and regulatory demands, particularly regarding system integrity, lifecycle management, and cybersecurity. The study employs a structured risk analysis across four dimensions—technical, operational, legal, and user-related—to assess the applicability of automotive standards to stationary fire alarm systems. Findings reveal significant overlaps in key risk areas such as failed updates, compatibility issues, cyberattacks, and liability. Consequently, the layered approach of ISO 24089, which distinguishes between organizational and technical update processes, offers a suitable framework for fire alarm technology. Similarly, the lifecycle-oriented risk management approach of ISO/SAE 21434 can be adapted to enhance cybersecurity resilience in building-integrated systems. Despite these promising parallels, domain-specific constraints—such as stricter certification requirements and differing network infrastructures—must be taken into account. The qualitative methodology used provides a valuable conceptual basis for standard transferability but lacks empirical validation. Future research should therefore include practical trials and extend the approach to related domains such as industrial automation (e.g., ISO/IEC 62443). Overall, the study demonstrates that cross-domain knowledge transfer can strengthen the security, maintainability, and regulatory robustness of networked fire alarm systems, thereby contributing to the broader discourse on resilient digital infrastructures.

Abstract:
Advanced Driver Assistance Systems (ADAS) are vulnerable to physical adversarial attacks in the infrared (IR) spectrum, which remain invisible to human drivers yet are captured by vehicle cameras. This paper presents a two-stage architecture for detecting and correcting IR adversarial noise when applied on traffic signs 1 . The first stage employs classical image processing techniques to extract the region of interest (ROI) and detect anomalous patches using color segmentation in the HSV space. The second stage leverages a Context Encoder, trained adversarially, to semantically inpaint the masked area based on surrounding visual context. A convolutional neural network (CNN) classifier is then used to accurately recognize the reconstructed sign. The proposed system was tested on a custom- augmented GTSRB dataset including synthetically corrupted IR patches. Results show that our method achieves a PSNR of 26.64 dB and SSIM of 0.81 for reconstruction, and classification accuracy of 95.85% for the full system, significantly improving robustness compared to baselines. This approach enables real- time, software-based mitigation of IR adversarial attacks without requiring hardware changes.

Abstract:
Autonomous driving involves the challenge of assessing and managing risks in complex environments. One approach for addressing this is the use of human-readable explanations of risk-related tasks to support justification of driving behavior. In this paper, we introduce VISION-RISK, a vision-language model (VLM) designed for risk assessment in autonomous driving using a lightweight architecture, optimized for deployment on edge devices. To train the model, we developed a custom dataset combining real-world driving scenarios from Honda Driving Dataset and extreme high-risk cases from Car Crash Dataset, augmented with synthetic annotations using Dolphins and refined via DeepSeek-V3. VISION-RISK stands out through three key characteristics: the integration of danger level classification with natural language explanation generation, a lightweight architecture optimized for deployment on resource-constrained devices, and a focus on safety through risk assessment to support trust in autonomous driving.

Abstract:
We present a method to analyze the extrapolation capacity of convolutional neural networks, for classification tasks, in the context of model-in-production performance expectations. More precisely, we investigate the calibration capacities of the quantized int8 and float8 models after training, assessing the suitability of deployment to production, by taking into consider- ation conformal prediction metrics alongside the well-established classification metrics. We prove that conformal metrics, such as APS[1], RAPS[2] and ECRC [3] provide additional insights into the prediction capabilities of the models, and should, as such, be used in conjunction with classical machine learning metrics as an additional evaluation tool when determining the overall reliability of a model.

Abstract:
One of the most critical factors influencing success in Formula 1 racing is the ability to make strategic decisions at the optimal moment. These decisions—such as pit stop timing and tire selection—are deeply tied to lap time dynamics. In this study, we propose a Long Short-Term Memory (LSTM)-based time series model for predicting the next lap time of a driver using real-world multidimensional race data. The model is trained on data from the 2021–2023 seasons and evaluated on the 2024 season. A comprehensive data preprocessing and feature engineering pipeline is introduced, integrating race-specific variables with driver and team characteristics to enhance predictive performance. The proposed architecture effectively captures both sequential dependencies and contextual information, leading to more accurate and realistic lap time forecasts. Beyond its predictive capability, the model offers a promising foundation for real-time race assistants and decision-support systems. Future enhancements may include driver-personalized architectures, dynamic race condition modeling, and integrated strategic recommendation systems.

Abstract:
Accurate forecasting of electricity demand plays a crucial role in modern energy systems, enabling efficient planning, maintaining grid stability, and optimising resource allocation amid increasing consumption and variable demand patterns. This paper presents HeySF (Hybrid EnergY Series Forecasting), a modular method for forecasting electricity consumption that integrates statistical models (SARIMA) and machine learning techniques (XGBoost, Random Forest, MLP, LSTM and PSO- optimised SVR). The approach is validated on real-world consumption data from the PJM Interconnection for daily energy consumption forecasting. The paper evaluates both individual predictors and ensemble schemes using standard error metrics, and includes a cost-aware analysis of training and inference latency measured as wall-clock time.

Abstract:
Centralized processing of data from IoT devices can generate privacy issues, as sensitive information, such as medical information, would be transmitted to a less secure cloud system. In classification problems, one solution would be to use federated learning (FL) to generate a global model from locally trained classifiers without transferring raw data through the network. There is always a trade-off between privacy and performance that should be reviewed when designing an IoT system. In this paper, we conducted experiments in two use cases (myocardial infarction detection and smoke detection) to demonstrate the applicability of FL in various domains. Using multiple federated learning techniques, we show that the small negative impact on accuracy is reasonable considering the privacy benefit of FL. The experiments show that with an appropriate aggregation scheme, FL obtains accuracy similar to that of the centralized approach and outperforms the non-FL distributed computation.

Abstract:
Renewable Energy Communities (RECs) empower prosumers to integrate renewable energy sources at the network edge. However, the intermittent generation of energy can produce sudden spikes that risk line overloads and voltage excursions. Optimizers can improve local grid management, and Particle Swarm Optimization (PSO) is a popular approach for its ability to handle non-linear dynamics, however it struggles with high-dimensional contexts. In this paper, we propose a distributed, fully on-chain Particle Swarm Optimization (PSO) algorithm, for transparent optimization within RECs. Each prosumer deploys a Node smart contract that runs a local PSO instance, enhanced for the energy domain, which considers individual energy profiles and custom penalties to compute a local cost. Then, a lightweight Global Coordinator contract aggregates each node’s best positions, weighted by their flexibility band, to compute a system wide plan which every prosumers tries to follow in an iterative process. We evaluate the solution using real consumption data in scenarios with up to 500 prosumers, showing that the solution converges and reduces the cost of each prosumer, while increasing the renewable energy usage in the community.

Abstract:
This paper proposes a decentralized framework for managing energy transfers between interconnected microgrids (MGs) and the main grid, using smart contracts implemented on a blockchain infrastructure. Using a predictive control mechanism (MPC) to estimate local consumption and production, the Distributed System Operator (DSO) can detect energy imbalances at the MG level and initiate the process of redistributing energy between MGs to cover the imbalances. Microgrids participate in the trading process through a coordinated notification system, avoiding allocation conflicts. The selection of trading partners is based on transfer scores in a greedy manner, and energy exchanges are tokenized through the ERC-20 standard. Experimental results indicate that, following the trading process, MGs reach an energy balance, with stability levels ranging between 85.85% and 90.13%. On average, 94.3% of the energy demand is covered locally, and only 5.6% comes from imports from the main grid.

Abstract:
European microelectronics value chain for highly automated driving and green mobility (EuroDrives), within the European Important Project of Common European Interest on Microelectronics and Connectivity (IPCEI ME/CT), advances video-based perception and automated driving through the development of next-generation driver assistance controllers and systems. The project focuses on AI technologies for self-driving vehicles, emphasizing three main directions: advanced multi-modal 3D perception and fusion methods, AI for pushing the boundaries of automated driving and embedded AI optimizations for low-power hardware. The system integrates centralized architecture and full-surround computer vision to enable reliable driving and parking automation. The session will give an update on the following activities: 1) Metric 3D video-only offline perception 2) Online maps generation for automated driving 3) Painted symbols detection and the use of foundation models 4) Precise landmarks 3D positioning with multi-view optimization techniques 5) Embedded AI: bridging the gap between cloud AI models and automotive SOCs

Abstract:

Abstract:

Abstract:
This paper presents the development of an AI-based anatomical landmark recognition framework designed for robot- assisted dynamic dental implant surgery. The proposed framework integrates a YOLOv11 deep learning model with a RealSense D405 RGB-D (Red, Green, Blue and Depth) camera, mounted on a collaborative robot, to detect predefined dental landmarks in real time. These 2D detections are projected into 3D space using depth information and subsequently registered to a preoperative STL model derived from CBCT (Cone Beam Computed Tomography) scans using Unity. The implant position, defined in the planning phase relative to anatomical landmarks, is localized intraoperatively without the use of fiducial markers and projected on the STL model of the mandible/maxilla.

Abstract:

Abstract:

Abstract:

Abstract:
In this paper, we introduce CTF-PWN100, a dataset and automated framework designed to evaluate the performance of large language models (LLMs) in solving Capture-The-Flag (CTF) binary exploitation challenges. We specifically assess the capabilities of DeepSeek-R1, Gemini-2.5-Pro (Preview), and Gemini-2.5-Flash (Preview) on a balanced and diverse dataset consisting of 100 binary exploitation problems covering multiple vulnerability types, including Buffer Overflow, Format String, Heap Overflow, Use-After-Free, and Miscellaneous. Using an automated evaluation framework integrating decompilation, code filtering, and dynamic feedback from Python-generated exploit scripts, the three models achieved success rates of 51%, 66%, and 46%, respectively. The results indicate that reasoning-oriented LLMs demonstrate promising potential in addressing complex cybersecurity tasks. However, limitations persist regarding vulnerability identification and the development of sophisticated exploit scripts.

Abstract:
Software Composition Analysis (SCA) is an important part of the reverse engineering process, useful for both antimalware research and vulnerabilities scanning. While traditional SCA techniques rely on Software Bill of Materials (SBOM) and strings analysis, we propose a new technique, based on binary code analysis on Rust programs. The binary code is analyzed through disassembly and OpCode n-grams extraction. The n-grams are then used to tag Rust crates using multiple methods and thresholds. The experimental evaluation shows the robustness of the proposed technique, being able to distinguish between 150 popular crates with over 87% accuracy and 48% F 1 score (micro-averaged) on the validation data.

Abstract:
This paper presents the design and implementation of a stealthy dynamic analysis environment for Linux, named Rootalyx. The system addresses the challenge of evasive mal- ware that can detect and circumvent traditional monitoring tools. Its core component is a kernel module that uses the ftrace subsystem to instrument system calls, integrated with rootkit- inspired hiding techniques to ensure a covert operational profile. The collected behavioral data, including system call arguments, is processed by a hierarchical hybrid model combining the LSTM and Transformer architectures for automated malware classification. Experimental results show that the proposed system introduces lower performance overhead compared to traditional tools for monitoring like strace and successfully evades detection by common security scanners. The machine learning model achieves a classification accuracy of 80.34% on the extracted data, validating the viability of kernel-level stealth monitoring for robust malware analysis.

Abstract:
Security operations centers (SOCs) are increasingly deploying large language models (LLMs) to assist human analysts; however, the reliability of off-the-shelf LLMs for malware analysis remains uncertain. This paper investigates whether state-of-the-art LLMs can serve as “virtual malware analysts” without task-specific fine-tuning. We introduce ArA (Artificial Analyst), a novel system that combines dynamic malware sandboxing with an LLM-based reasoning engine. ArA executes programs in a controlled environment, translates the observed behaviour into natural-language reports, and prompts an LLM to classify the behaviour as malicious or benign. We evaluated 12 modern LLMs—including GPT-4, its distilled variants, open-source models (1B–27B parameters), and a security-tuned model—on a dataset comprising 712 real malware samples and 304 benign programs. GPT-4 achieved the highest accuracy at 95% (macro F1 = 0.93), correctly identifying nearly all malware samples and rarely mislabeling benign software. Smaller open-source models demonstrated mixed performance; some attained perfect malware recall but exhibited higher false-positive rates on benign data, while the security-tuned 8B model performed below general-purpose models. We analyse each model’s strengths, common errors (e.g., false-positive biases), and the usage of tentative (“suspicious”) versus definitive labels. Our findings demonstrate that advanced LLMs can accurately distinguish malware behaviours in zero- shot settings, offering immediate value for SOC triage. However, we also highlight limitations, including insufficient coverage of behaviour variations and limited explanation quality, and discuss future steps to enhance LLM-based malware analysis for practical deployment.

Abstract:
Operating System (OS) deduction is a dangerous technique, aimed at devices during the scouting period preceding a targeted attack. It aims to identify certain vulnerabilities known to affect an OS. Nmap stands out as the leading tool for this job, offering both popularity and reliability. It uses the ambiguities in different protocols such as TCP, IP and ICMP, with different OS vendors choosing different values. This paper compares three machine learning classifiers, neural networks (NNs), random forest (RF), and support vector machines (SVM), to determine the most appropriate and precise approach that could be used for this purpose. A dataset containing both normal and OS fingerprinting traffic was generated from a relevant collection of modern devices ranging from iPhones, laptops, desktops, to Internet of Things (IoT) devices. However, the research also adopted well-known and publicly available datasets for comparison. As shown by the results, neural networks have bigger difficulty in detecting Nmap traffic, while SVMs have higher false positives. Random forests outperform both, achieving over 99% accuracy (even with just 10 trees), low false positives, and strong performance with fewer features, making it the most effective and robust choice.

Abstract:
We introduce a fully decentralized strategy enabling teams of non-holonomic mobile robots to autonomously form and maintain stable, configurable line-shaped formations during navigation. Central to our approach is a lightweight communication backbone implemented as an approximate minimum spanning tree (MST), constructed efficiently in two stages. First, robots use local geometric criteria to rapidly form initial spanning fragments, which are then incrementally merged through a decentralized, multi-scale Boruvka-inspired process. Leveraging this structure, we extend a potential field–based formation controller to incorporate local alignment behaviors, allowing each robot to dynamically adjust its orientation and inter-robot spacing according to user-defined configurations. This framework ensures robust and scalable formation control without relying on global knowledge or predefined network topologies. Extensive simulations confirm the strategy’s effectiveness, demonstrating rapid convergence, adaptability to evolving formation configurations, and guaranteed asymptotic stability.

Abstract:
The growing deployment of Automated Guided Vehicles (AGVs) in industrial and logistical environments demands efficient and scalable collision detection mechanisms to ensure safe and uninterrupted operation. Traditional CPU-based methods struggle to meet the real-time performance requirements of large-scale, dynamic AGV systems. This paper presents a parallel collision detection algorithm specifically optimized for execution on Graphics Processing Units (GPUs). By leveraging the inherent parallelism of GPU architectures, our approach significantly accelerates proximity checks and collision resolution among multiple AGVs operating in complex environments. We explore data-parallel strategies, including spatial partitioning and bounding volume hierarchies, to manage computational load and reduce latency. Performance evaluations conducted on AGV fleets show substantial speed-up — up to 30x compared to OpenMP- based CPU implementations — while maintaining high detection accuracy. The results highlight the potential of GPU-accelerated collision detection to enhance responsiveness and safety of next- generation AGV systems.

Abstract:
In crowded expositional environments such as museums or exhibitions, mobile robots face significant challenges in navigating towards groups of visitors to deliver specialized guidance. This paper explores a practical approach for estimating crowd density using Bluetooth Low Energy (BLE) signals detected by an ESP32-based scanner placed on a guidance robot. Since many people carry devices like smartphones or smartwatches that emit BLE signals, the system uses these signals to estimate how many visitors are nearby and in which direction. The scanner rotates through four fixed angles with the help of a servo motor and shielding, allowing it to measure signal strength (RSSI) from different directions. Based on these measurements, the robot can move towards areas where more people are gathered. The system was tested in various setups, and the results—compared with measurements from two mobile devices—indicate that it can provide consistent data. While relatively simple, this method could serve as a useful and privacy-friendly alternative to camera-based systems for guiding robots in indoor environments.

Abstract:
Multi-Agent Pathfinding (MAPF) ensures collision-free motion for teams of autonomous robots, yet practical applications also demand intelligent distribution of tasks across those robots. We present a fully integrated framework that couples optimal path planning with dependency-aware task allocation and validated from graph-based models to simulation and real- world robots. First, we develop scalable MAPF algorithms in a graph abstraction and study their complexity, makespan and sum-of-costs behaviour. These algorithms are then embedded in a ROS 2 + Gazebo environment that incorporates realistic kinematics, network delays and partial observability; Cartographer provides on-line occupancy grids while the Nav2 stack supplies local/global planners. To elevate the system from ’just moving’ to ’moving with purpose’, we formulate task assignment with precedence constraints as a combinatorial optimization problem and solve it using an Ant Colony Optimization (ACO) meta-heuristic that minimizes total completion time and idle time of the robot. The same software stack is ported to differential drive robots equipped with 2-D LiDAR and Raspberry Pi controllers. By demonstrating end-to-end transfer from theory to real robots, this work establishes a reproducible pathway for deploying cooperative, task-driven multi-robot systems in manufacturing, logistics, and smart-city scenarios.

Abstract:
The classification of crime-related news articles poses unique challenges due to overlapping labels, subtle semantic cues, and limited annotated data. This paper proposes a hybrid semantic-symbolic classification framework that combines state- of-the-art sentence-transformer embeddings with domain-specific symbolic reasoning. We benchmark multiple transformer-based models, including MiniLM, BERT, and the multilingual E5-Large model, for embedding Italian crime news, and evaluate traditional classifiers (Support Vector Machines, Logistic Regression) as well as deep learning models (Convolutional Neural Networks). To enhance explainability and performance, we integrate Basic Formal Ontology (BFO)-compliant mappings using a symbolic post-processing layer based on predicate logic and ontological relationships. Experiments on a curated Italian crime news dataset demonstrate that our hybrid approach significantly improves classification accuracy and macro-F1 scores compared to zero-shot and purely statistical baselines. In particular, a CNN combined with E5-Large embeddings and clustering achieves the highest performance, with a macro-F1 score of 74 0%. The results suggest that integrating structured domain knowledge into neural pipelines enhances model transparency and robustness for multi- label crime classification.

Abstract:
Image security plays a crucial role in various domains where sensitive information is transmitted, such as telemedicine, satellite imagery transfer, and military communications. Effective image encryption requires strong security, secure network transmission, and low computational cost. Traditional encryption methods provide robust security and reliable transmission but are computationally expensive. Chaos-based cryptosystems offer lower computational costs but generally lack sufficient network security. Experimental neural network- based encryption achieves high processing speed but introduces overhead due to model training and is limited by its dependence on specific datasets, making it non-adaptive. In this work, we propose a novel image encryption scheme combining chaotic theory, elliptic curves, and static S-boxes. The method initially employs the CCS chaotic map to generate a master key sequence used for encrypting the plain image. These static keys are further utilized in a symmetric process involving secret key derivation based on scalar operations over elliptic curves and S-box transformations. Experimental analysis demonstrates that the proposed scheme offers strong resistance to side-channel and statistical attacks, including differential analysis, outperforming several existing methods in terms of security. The main limitation lies in the increased computational cost due to elliptic curve operations used in the key derivation process.

Abstract:
he absence of debugging symbols in stripped binaries, particularly descriptive function names, hinders reverse engineering and security analysis. This paper introduces a structured pipeline for automated function name inference, combining Transformer models, calibrated confidence estimation, and context-aware refinement using large language models pretrained on code. In the first stage, a specialized Transformer model is trained on normalized assembly instructions to predict initial function names, along with statistical indicators. These indicators are then mapped to reliable confidence scores using a logistic regression calibration module. In the final stage, low- confidence functions are reprocessed using fine-tuned, pretrained code language models, which leverage the context derived from high-confidence caller and callee relationships. The method is validated on a dataset of approximately 21.3 million functions extracted from 6,824 Windows executables compiled with various MSVC versions and architectures. The results demonstrate that our approach supports the practical recovery of meaningful function identifiers and enhances malware analysis tasks.

Abstract:
In order to address the necessity of invisible image watermarking, this paper proposes a robust and imperceptible method for embedding data into an image. It makes use of a variable size DFT kernel in order to subdivide the original image, and in each subdivision several concentric high-frequency bands are chosen in order to encode a bit of the watermark image.

Abstract:
In recent years, artificial intelligence has played a pivotal role in advancing medical imaging, particularly in the early detection of thoracic diseases through chest X-rays. Latest efforts have been successful in proposing deep learning solutions for accurate classification of different pathologies. The objectives have been refocused towards improving the detection of different signs of disease. We propose an original deep learning model for affected area detection. The model is based on the YOLOv12 object detection architecture, which was enhanced specifically for medical use cases. The improvements include attention mechanisms, deformable convolutions, and feature aggregation modules such as FPN and PANet. We created a custom dataset, mainly a combination of two public chest X-ray datasets, which were augmented to train our model. Experimental results demonstrate that the proposed model achieves a mAP@0.5 of 0.374 and a Recall of 0.519, outperforming existing models, including baseline YOLOv12, YOLOv11, YOLOv8, and Faster R-CNN on these key metrics. Despite the architectural enhancements, inference time remained competitive at 6.3 ms per image. The model shows high reliability in identifying four key thoracic pathologies: cardiomegaly, aortic enlargement, pulmonary nodules/masses, and pleural effusion. These findings suggest that optimized YOLOv12-based models can serve as effective diagnostic support tools in clinical and remote healthcare settings.

Abstract:
Automated evaluation of rehabilitation exercises is essential for supporting home therapy and long-term recovery for patients with physical impairments. This study investigates several deep learning-based approaches for the assessment of gesture correctness, comparing two original Long Short-Term Memory based architectures with three deep convolutional models previously proposed in the literature for time series classification. While most existing approaches use separate models for individual exercises, we employ a single, general model for all gestures. The experimental evaluation was performed using two different evaluation strategies: a standard split, with mixed subjects in training and testing and a more realistic leaveone-subject-out (LOSO) procedure, that performs the tests on unseen individuals. We employed two rehabilitation datasets in our experiments and the results show that LSTM-based models generally outperform convolutional architectures, particularly for the LOSO strategy.

Abstract:
This study presents a deep learning framework for automatic segmentation of orbital walls in cranial CT scans using U-Net-based architectures under multiple training regimes. Three experiments were conducted: (i) a supervised baseline trained on nine expert-annotated volumes from the public CranealCT dataset; (ii) a semi-supervised model with 50 pseudo-labeled volumes; and (iii) a final model trained on 39 heterogeneous volumes, combining public and proprietary data with expert annotations. Preprocessing included intensity normalization, CLAHE, Gaussian smoothing, and window standardization. Training used Dice and Binary Cross-Entropy loss with strong augmentation. The semi-supervised model outperformed the baseline (Dice 0.968 vs. 0.937). On the proprietary test set, the final model achieved Dice 0.9193, HD95 2.53 mm, ASD 0.87 mm, accuracy 0.9985, precision 0.8978, and recall 0.9763, demonstrating robustness to imaging variability and suitability for clinical use in orbital trauma reconstruction.

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:
Electroencephalography (EEG) remains one of the most widely used non-invasive methods for investigating brain function, due to its high temporal richness and accessibility. Still, the complex and noisy nature of EEG signals makes decoding neural states a challenging task, particularly when using traditional signal processing or shallow machine learning approaches. In recent years, deep learning models, such as Transformers, have emerged as a powerful alternative to classical machine learning methods, thanks to their capacity for modeling long-range dependencies and learning contextual representations. In this paper, we explore the application of Transformer architectures to EEG data for the purpose of binary and three-class classification in a visual perception task. To evaluate model performance, we compare the Transformer against four other models: XGBoost, Random Forest, a simple Multi-Layer Perceptron and a Long Short-Term Memory network. Our findings suggest that attention-based models are capable of capturing informative temporal dependencies in EEG signals, with preprocessing playing a critical role in optimizing representation quality. This work highlights the potential of Transformer models for non-invasive neural decoding and contributes to the development of more effective tools for cognitive neuroscience and brain-computer interfaces.