Abstract:
This paper proposes an artificial intelligence model designed to identify the positions of the fingers on the hands, assigning them corresponding musical notes. The creation of this model involved the development of a custom dataset comprising various hand configurations, each specifying how a person mimics pressing a piano key. To expand the volume of data and improve its variety, data augmentation was used, a technique that helps the model perform more efficiently under various conditions of use. The data processing included normalization of the coordinates relative to the wrist, so that aspects such as the size, color or shape of the hands do not influence the performance of the model, and the actual position of the hands is independent of the distance from the camera. Another innovative aspect of the development was the application of contrastive learning, whereby the model distinguishes between correct and incorrect hand positions, forming pairs of similar data and minimizing the distances between their coordinates. The results show that the model achieved a high performance for hand gesture classification with an accuracy of 98.89%.

Abstract:
In this paper we evaluate the performance of four machine learning algorithms (i.e., Gradient Boosting, Random Forest, Decision Trees and K-Nearest Neighbors) for classifying emotional states and detecting stress using physiological data collected from chest and wrist wearable devices. We have performed a comprehensive analysis involving feature extraction and selection from physiological data, followed by hyperparameter tuning for each algorithm. The performance of the resulting models has been assessed using the accuracy, recall, F1 score, and loss metrics, alongside confusion matrices to provide insights into the effectiveness of classification. The paper analyzes how well each algorithm classifies emotional states, highlighting their strengths and weaknesses.

Abstract:
Human action recognition (HAR) systems need to process large volumes of data posing several challenges including, but not limited to, accurately identifying the actions and classifying them in near real time. Most of the state of the art solutions rely on machine learning techniques where a substantial amount of time is required both for the training and the inference phases. Existing solutions try to reduce the size of the data that is processed by the HAR classifier by constructing a spatiotemporal region of interest (ST-ROI) based on skeleton data, such as the model-based multimodal network (MMNet). However, the performance of these solutions depends on the accuracy of the skeleton data. This paper presents a novel solution to reduce the size of the data that is processed during classification by constructing an ST-ROI using regions identified as motion salient. The results show that the classification accuracy of the proposed solution is comparable to existing solutions such as MMNet but with a training time that is drastically lower.

Abstract:
The acceptance of remote care solutions is steadily advancing, especially in assuring accessible healthcare services, monitoring vital signs and offering support for older adults. New Internet of Things (IoT) systems have been developed to tackle the healthcare professionals needs to track the health progress elders at appropriate intervals with the goal of assessing their condition remotely and planning timely interventions. This paper proposes a platform designed to meet healthcare needs of seniors and medical personnel through remote patient monitoring of activities (e.g. effort, sleep) and health parameters (e.g. blood glucose level, blood pressure, heart rate, Spo2, etc.). The platform takes advantage of the capabilities of modern wearable devices and smartphone technologies to capture live medical data and utilizes neural network algorithms to enable advanced data analysis to identify possible health problems. Including tailored access for care professionals, health analytics insights and integrated communication channels, the platform meets the current demands in the domain of remote healthcare services. The platform evaluation featured good results for the prediction algorithms on literature datasets and validated the feasibility of the overall remote monitoring infrastructure

Abstract:
Decision-making for automated driving remains a challenging task. For their integration into real platforms, these algorithms must guarantee passenger safety and comfort while ensuring interpretability and an appropriate computational time. To model and solve this decision-making problem, we have developed a novel approach called COR-MP (Conservation of Resources model for Maneuver Planning). This model is based on the Conservation of Resources theory, a psychological concept applied to human behavior. COR-MP is based on various driving parameters, such as comfort, safety, or energy, and provides in real-time a profit value that enables us to quantify the impact of a decision on the decision-maker. Our method has been tested and validated through closed-loop simulations using RTMaps middleware, and preliminary results have been obtained by testing COR-MP on a real vehicle

Abstract:
Autonomous mobile systems exploring unknown environments face the challenge of optimizing their routes to minimize time and distance traveled. This paper explores integrating SLAM algorithms with optimized routing techniques to address the Mixed Chinese Postman Problem (MCPP). The MCPP arises when a robot aims to efficiently explore an area by determining the most optimal path that covers all edges of a graph while minimizing traversal distances. A key contribution of this paper is introducing an easy-to-implement solution for the MCPP in the context of autonomous mobile systems. We discuss the integration of this solution within the SLAM framework, highlighting its compatibility and effectiveness in real-world scenarios.

Abstract:
This paper addresses the task of road scene classification by leveraging multi-task learning (MTL) to improve model performance and computational efficiency. We employ the BDD100K dataset, which provides diverse driving scenarios, and apply data augmentation and balancing techniques to mitigate class imbalances. Our methodology integrates a ResNet18 backbone pre-trained on ImageNet, extended with task-specific heads for scene type, time of day, and weather classification. Uniform Manifold Approximation and Projection (UMAP) is used for dimensionality reduction and visualization of feature embeddings. Experiments demonstrate that our MTL approach significantly enhances classification accuracy across tasks, with notable performance in time-of-day classification. The proposed model achieves robust results despite class imbalances, validated through confusion matrices and UMAP visualizations. Our findings highlight the benefits of MTL in road scene classification, offering insights for future research in autonomous driving and computer vision.

Abstract:
In this paper, we introduce a multitask learning architecture for simultaneous depth estimation and surface normals prediction, built upon a shared encoder framework. Our approach integrates well-established depth estimation models to reduce computational complexity while enhancing feature representation for both tasks. Through extensive experiments on benchmark datasets, we provide both quantitative and qualitative analyses of the results. Our evaluations demonstrate significant improvements in depth estimation accuracy and efficiency compared to other state-of-the-art methods, along with competitive results in surface normals prediction. Additionally, the proposed architecture offers a lightweight design for multitask learning, presenting a robust and versatile framework for future research and practical applications. These findings emphasize the potential of multitask learning to enhance the efficiency and performance of related computer vision tasks.

Abstract:
Anomaly detection in time series data is a critical challenge across various domains. Traditional methods typically focus on identifying anomalies in immediate subsequent steps, often underestimating the significance of temporal dynamics such as delay time and horizons of anomalies, which generally require extensive post-analysis. This paper introduces a novel approach for time series anomaly prediction, incorporating temporal information directly into the prediction results. We propose a new dataset specifically designed to evaluate this approach and conduct comprehensive experiments using several state-of-the-art methods. Our results demonstrate the efficacy of our approach in providing timely and accurate anomaly predictions, setting a new benchmark for future research in this field

Abstract:
This paper provides an algorithm for the Controller Area Network (CAN) protocol, which enables transmission of CAN single frames and CAN flexible data-rate frames in an automatic manner. The automotive industry is of key importance nowadays and, in this context, our research focuses on autonomous/ automatic diagnosis and predictive maintenance techniques, which can also be easily extended to other important areas such as mechanical engineering, robotics, avionics, healthcare and so on. Our study shows improvements of up to 14% for validating and interpreting the data with respect to the response time of the ECU, which is not negligible for real-life and industrial applications, where time efficiency in detecting malfunctions is a sensitive issue.

Abstract:
In the context of Industrial Internet of Things, the deployment of AI-based anomaly detection algorithms in resource-constrained environments is critical. This paper systematically evaluates the complexity and real-time performance of widely used machine learning models—Random Forests, Support Vector Machines, Linear Regression, XGBoost, Neural Networks, and Autoencoders. Through comprehensive theoretical analysis and empirical measurements on four diverse datasets, we identify the strengths and limitations of each model in terms of computational demands and practical feasibility. Our findings highlight that while complex models like Neural Networks and Autoencoders offer high accuracy on time series datasets, their deployment on edge devices necessitates significant optimization for real-time performance. In contrast, simpler models, though less resource-intensive, may fall short in critical applications. This study provides a decision matrix and structured evaluation framework to guide the selection of efficient anomaly detection models tailored to specific operational constraints. These contributions aim to assist developers and system administrators in achieving balanced AI deployments that optimize performance, security, explainability, and resource consumption.

Abstract:
n the current digital landscape, protecting networks against malicious activities is a critical challenge. Network Intrusion Detection Systems (NIDS) are vital as the first line of defense, continuously monitoring network traffic to detect and prevent potential attacks in real-time. As cyber-attacks get more complex and more similar to normal traffic, robust NIDS solutions have become more crucial than ever. This paper proposes an architecture for an anomaly-based NIDS that has a Multi-Layer Perceptron (MLP) as a binary classifier. The system employs the Cisco TRex generator to simulate network traffic, capturing and analyzing the data using tcpdump and Zeek, and lastly preprocessing it for the MLP using Python scripts. Three algorithms were evaluated for the classification task: Isolation Forest (IF), MLP, and Autoencoder, all implemented with TensorFlow and Keras. The models were trained and tested on two widely recognized datasets for anomaly detection, KDDCUP99 and UNSW-NB15. The experimental results show the superiority of the UNSW-NB15 dataset compared to the KDDCUP99 one in terms of complexity and its likeness to realworld traffic. Moreover, the results also prove that a simple Deep Learning (DL) algorithm such as the MLP can serve as an effective first-line defense against cyber threats. This study contributes to the ongoing development of more effective NIDS by exploring the application of machine learning techniques in anomaly detection, offering the potential for enhancing network security.

Abstract:
Wildlife on public roads presents a public safety problem as well as an environmental concern. The field of object detection within Computer Vision offers some approaches that can be used either in autonomous vehicles or within emergency braking systems for collision avoidance. This paper aims to study how a popular and efficient object detection model and a keypoint-based model compare against each other on an SoC specialized for Computer Vision tasks which is used in automotive systems. Insights into how these object detection models perform on resource-constrained hardware and into the importance of the overall approach are gained while also highlighting the advantages and disadvantages of each method and the hindrances of deploying such algorithms on hardware. While well-established models like YOLO outpace the implemented keypoint-based approach in both precision and speed, the total runtime of the two models is ultimately close, with the biggest takeaway being that the robust YOLO model also requires extensive post-processing, which is computationally expensive, especially on an embedded SoC.

Abstract:
As cities are becoming more crowded due to the population moving to urban areas, smart parking and management are essential elements of the transportation system. There are efforts in developing vision-based automated parking systems by detecting the ground markings accurately in real-time from multiple surround-view cameras. We are not interested in automated parking, but in collecting relevant parking occupancy statistics from the city streets. In this paper, we propose a novel method based on a single front-view camera and a convolutional neural network (CNN) for detecting the parking slots and classifying them based on occupancy. This information collected by the camera is sent to a central processing cloud and can serve as a basis for many Smart City parking management applications, by providing useful dynamic occupancy statistics.

Abstract:
Fault diagnosis is crucial for complex autonomous mobile systems, especially for modern-day autonomous driving (AD). Different actors, numerous use cases, and complex heterogeneous components motivate a fault diagnosis of the system and overall system integrity. AD systems are composed of many heterogeneous components, each with different functionality and possibly using a different algorithm (e.g., rule-based vs. AI components). In addition, these components are subject to the vehicle’s driving state and are highly dependent. This paper, therefore, faces this problem by presenting the concept of a modular fault diagnosis framework for AD systems. The concept suggests modular state monitoring and diagnosis elements, together with a state- and dependency-aware aggregation method. Our proposed classification scheme allows for the categorization of the fault diagnosis modules. The concept is implemented on AD shuttle buses and evaluated to demonstrate its capabilities.

Abstract:
We investigate various methods of neural network optimization with the purpose of reducing runtime and maintaining performance when deployed on embedded devices. The proposed approaches follow both improvements upon the general architecture of the deep neural models, as well as possibilities of increasing hardware utilization and employing scheduling and parallelization techniques. The presented solutions significantly improve the runtime and the behaviour of neural networks when deployed on embedded targets. The optimization strategies improve the runtime by approximately 3%-4% when applied on traditional Convolutional Neural Networks. Meanwhile, the analyzed Birds-Eye View Transformer benefited from an improvement ranging from 38% up to 83% depending on the utilized hardware and the quantization range. Thus, we stress the importance of taking advantage of the available hardware resources in order to reach the best possible results in terms of execution runtime

Abstract:
Time series are one of the most common data types in data mining, capturing sequential observations over long periods. Creating synthetic time series data involves generating a sequence of data points that maintain the statistical properties found in real-world datasets. This article addresses the generation of synthetic time series data from real univariate time series datasets. It presents two novel methodologies grounded in Seasonal-Trend decomposition using LOESS (STL). The first method involves estimating the probability distribution of the residual component and generating new residuals to substitute in the STL decomposition. The second method replaces the noise component with a randomized version and synthesizes seasonality using Long Short-Term Memory (LSTM) neural networks. We applied these methodologies to time series data representing the operating times of household appliances. The synthetic data produced were assessed both quantitatively and qualitatively using a range of statistical indicators. Moreover, we improved the classification performance on imbalanced real data with synthetic data.

Abstract:
Synthetic data generation plays a crucial role in various domains where access to real-world datasets is limited or restricted due to legal, ethical, or privacy concerns. Given the growing need for realistic data in analysis and research, we explore three distinct methods to capture real time series data’s temporal and distributional characteristics. We provide a comparative study of these methods by using a set of evaluation metrics, highlighting the strengths and benefits of each approach. This analysis offers multiple perspectives on the utility and applicability of the proposed synthetic data generation techniques. Additionally, we investigate the impact of synthetic data when forecasting time series, shedding light on its potential for enhancing predictive models.

Abstract:
Lung cancer detection remains a challenging task due to the complex nature of medical imaging and the shortage of specialized radiologists. This paper addresses the challenges of early detection of lung cancer using advanced machine learning techniques for CT image analysis. Given the limitations of current semantic segmentation methods and the lack of interpretability, we propose an innovative approach combining classification techniques with explanatory methods to enhance both precision and transparency. We developed a novel data preprocessing pipeline, a multi-stage CNN classifier training model, and a segmentation technique using concatenated Grad-CAM heatmaps. Our evaluations demonstrate significant improvements in tumour detection and model explainability, providing more effective support for radiologists. These advancements are reflected in increased segmentation accuracy, as evidenced by comparative analyses and qualitative assessments.

Abstract:
This paper presents a novel clustering algorithm, named Drop Ripple Counter, designed to address limitations inherent in traditional clustering techniques such as K-means and DBSCAN. The proposed algorithm was evaluated against these clustering techniques using a diverse set of datasets. The performance comparison focused on two key clustering validation metrics: Adjusted Rand Index and Adjusted Mutual Information. Experimental results demonstrate that our algorithm is on par or outperforms K-means and DBSCAN across multiple datasets in terms of the previously mentioned metrics, indicating superior performance in identifying underlying data structures and robustness to noise. These findings highlight the potential of our approach as a more effective alternative for a wide range of clustering applications

Abstract:
Semantic and instance segmentation have seen significant advancements, yet their application to aerial imagery faces numerous challenges. The high complexity of aerial images which often capture extensive scenes with numerous objects and varying object scales, differ considerably from the simpler natural images that the visual deep learning models are usually trained on. Additionally, the scarcity of annotated datasets for pixel-level semantic and instance segmentation in aerial imagery, hinders the development of specialized systems. However, recent progress in foundational models trained on vast, diverse datasets has improved generalization capabilities, making them robust across various scenarios. These models can be applied directly or adapted with minimal annotated data, significantly reducing the need for extensive labeled datasets. Even though foundation models are consecrated for their high generality, the results on specific scenarios, such as aerial images, are not as good as expected. In this paper, we provide an automatic panoptic segmentation and annotation system adapted to the characteristics of the drone vision images based on the visual foundation models Grounding DINO, SAM (Segment Anything Model) and the visual language model CLIP. We propose an algorithm that combines the foundation models and includes processing techniques to improve the performance on aerial scenes, eliminating the need for pixel-level annotated training data. The algorithm can segment objects using a fixed number of semantic classes or segment new ones through a text input by the user. We also provide a demo web application to test our solution on images from the UAVid dataset.

Abstract:
Recent advancements in deep learning algorithms have revolutionised computer vision, significantly impacting precision agriculture and farming automation. Object detection and tracking techniques, supported by drones equipped with vision sensors, play pivotal roles in monitoring and enhancing agricultural processes. This study focuses on improving object detection for precision viticulture, addressing challenges such as the small size of grape bunches and drone-induced distortions. By integrating the Coordinate Attention mechanism into the GELAN network, we enhance the detection of grapes in high-resolution images. Additionally, the study evaluates three methods for stabilising grape tracking, with ORB feature matching proving superior. Overall, these advancements contribute to automated crop monitoring and increased agricultural productivity through robust computer vision techniques. The findings underscore the significance of tailored approaches in adapting object detection technologies to specific precision agriculture applications

Abstract:
UAV computer vision in precision agriculture can enhance efficiency and reduce environmental impact compared to traditional techniques. However, the computational limitations and power constraints of UAVs hinder their performance, especially for real-time deep learning tasks. Compressed neural network models offer a solution, yet fixed compression levels may result in unacceptable accuracy loss. This paper proposes a novel context-aware approach for energy-efficient, real-time on-UAV weed detection. By incorporating contextual factors such as brightness, saturation, contrast, and vegetation indices, our method estimates input image difficulty and dynamically selects the optimal neural network compression level. Leveraging slimmable neural networks, which enable training a single model with various widths during inference, our approach ensures the best accuracy-resource consumption trade-off. Experimental results with two different slimmable networks architectures showed that our approach can lead to up to 35% less computations while also increasing the quality of the inference with up to 2%. This significantly improves the practicality and performance of UAV-based weed detection systems.

Abstract:
Urban development is advancing at an exceedingly rapid pace, significantly complicating the task of updating reference maps and requiring considerable effort in terms of actualization and monitoring. To address this challenge, we propose an automated mechanism for detecting changes in maps, specifically at the building level. Our method integrates cadastral maps with low-altitude aerial imagery, identifying discrepancies between the reference maps and the perceived images. We use captured images from standard Unmanned Aerial Vehicles and derive the cadastral maps from OpenStreetMap. The change detection process employs a mechanism that localizes the drone’s position by aligning the perceived scene with the reference map. We utilize established foundation models, such as Grounded SAM and Depth Anything, for image-level perception based on semantic segmentation and depth estimation, capitalizing on their robustness and generalization capabilities. Subsequently, we construct a bird’s-eye view representation that mirrors the reference map and generate a set of discriminative features. The obtained features are utilized to formulate hypotheses as rigid transformations, which are tested by projecting the image onto the reference map and validated using the Intersection over Union (IoU) metric. Once the perceived image is aligned with the reference map, we obtain the global GPS position and camera orientation, implicitly determining the global location of the buildings. After achieving accurate localization, the focus shifts to identifying differences between the perceived image and the reference map. In terms of experimental results, we tested the method in an aerial environment using a subset of the UAVid dataset, considering significant GPS disturbances, and achieving high precision in localization regarding both the position and the camera orientation.

Abstract:
This paper focuses on adapting the functionalities of the FastPitch model to the Romanian language; extending the set of speakers from one to eighteen; synthesising speech using an anonymous identity; and replicating the identities of new, unseen speakers. During this work, the effects of various configurations and training strategies were tested and discussed, along with their advantages and weaknesses. Finally, we settled on a new configuration, built on top of the FastPitch architecture, capable of producing natural speech synthesis, for both known (identities from the training dataset) and unknown (identities learnt through short reference samples) speakers. The anonymous speaker can be used for text-to-speech synthesis, if one wants to cancel out the identity information while keeping the semantic content whole and clear. At last, we discussed possible limitations of our work, which will form the basis for future investigations and advancements

Abstract:
Speech recognition models have an important role in improving efficiency and accessibility across various industries. Due to its advantages and benefits, this technology has transformed sectors such as healthcare, automotive, and customer service. However, the speech recognition in low-resource languages such as Romanian remains relatively underdeveloped due to the lack of data for training deep learning models. For this reason it cannot be utilized in the healthcare industry where Romanian specialized terms are unknown to the speech-to-text models. To address this problem, we propose to fine-tune large speech recognition models on a specialized dataset with Romanian audio and medical discourse. We obtain this dataset in a semi-automatic manner by leveraging speech recognition and large language models. By adapting state-of-the-art fine-tuning approaches that are used for LLMs, we improve the performance of the model on Romanian medical discourse by 19.9% in substitution errors, 28% in deletion errors and 9% in word error rate. This approach paves the way for speech recognition technology usage across many Romanian industries.

Abstract:
This paper addresses the challenge of adapting a multilingual language model to an unseen language by using translations to augment the training dataset. We evaluate how fine-tuning a model with Hungarian data obtained through translations impacts performance on several classification tasks. Our experiments demonstrate that translating recipes into Hungarian and using them for training significantly improves the model’s accuracy, with the best results achieved using only 35% of the total Hungarian recipes available. Additionally, fine-tuning on multilingual data before incorporating Hungarian leads to better performance compared to direct fine-tuning on Hungarian alone. Our approach has an overall increase in accuracy of 9.37% for three classification tasks. These findings highlight the effectiveness of using translations and multilingual models to enhance performance in low-resource languages in particular, where collecting labeled data for training is challenging and expensive.

Abstract:
This paper presents a novel approach to improve cooking experiences by integrating advanced natural language processing and machine learning techniques within the food industry. The research focuses on developing a predictive model that optimally predicts cooking settings for nutrient categorization, cooked product aspect and moisture reduction level. By leveraging NLP techniques, the model aims to interpret and classify recipes across multiple languages, enabling smart ovens to adjust settings autonomously. Joint learning is employed to train a transformer-based model, by combining the losses of the three tasks. Challenges such as multilingual data variability, dataset imbalance and model interpretability are addressed through different methodologies including data augmentation, feature engineering and Explainable AI techniques. The outcomes contribute to improve cooking processes, reducing errors and enhancing user trust in AI-driven culinary solutions. The results indicate that Translate-Train-All translation method, where the training set is balanced across languages, significantly improves the model’s performance, by nearly 1.00%, especially for low resource languages. Also, adding the additional genre feature enhances the model’s knowledge and improves its predictive capabilities by almost 1.00%

Abstract:
Detecting deception through automated tools involves computational techniques to identify falsehoods in text, which is crucial for evaluating the truthfulness of testimonies in the legal field. The present research explores deception detection in Romanian texts using various methods. Initially, ANNs with different text representations, such as TF-IDF, LSA, and LIWC features were employed. These approaches were compared with BERT-based models on augmented data, assessing improvements in classification performance. Additionally, CNNs with embedding representations were used to examine the impact of context on detection accuracy. To gain deeper insights, LIME (Local Interpretable Model-agnostic Explanations) was applied to visualise the importance of specific words and phrases in court statements. The present work is particularly novel for Romanian, as research on textual perjury detection in this language is limited. The obtained results are competitive with existing literature, highlighting the potential of the proposed approaches and impacting the fields of forensic linguistics and natural language processing, especially given the lack of research in non-English languages.

Abstract:
Modern automotive systems are starting to incorporate complex functional software features. With an increase in system complexity comes different requirements for security. In this paper, we propose a design for secure logging that uses the functionalities of an automotive Trusted Platform Module. Our logging solution guarantees log integrity, authenticity, nonrepudiation and tamper-proofness, while assuming a threat actor that has physical access to the storage medium of the protected logs. An experimental performance assessment was conducted to measure the performance of the Trusted Platform Module functions used in logging with different cryptographic and hashing algorithms

Abstract:
A realistic assessment for the cybersecurity posture of an organization involves a red team that acts as an attacker, using offensive security tools to test the existing defenses. We propose MEPHISTO, a red team tool that emulates a malware infrastructure, targeting cloud containers. The tool simulates an entire cybersecurity kill chain, including weaponization, exploitation and the command and control. The experimental results show that MEPHISTO can be used in a realistic attack simulation, featuring stealth communication with the command and control server and security solutions evasion.

Abstract:
This paper outlines an analysis method for automatically identifying the required capabilities for the system calls defined by the Linux Kernel. Integrated in the LLVM compiler infrastructure, the static and path sensitive approach used aims to generate a complete set of logical equations, expressed in terms of the syscalls‘ argument configurations, that describe the relation between system calls and capabilities. In order to achieve the desired syscall to capability mapping a four stage analysis pipeline is proposed, each stage complimenting one of the abstraction layers used by the LLVM compiler. First, all possible execution paths, in terms of functions, are identified by the first stage, which is followed by the identification of the control flow blocks leading to calls to ensuing functions in the chain. Once all such paths are known, the third stage is tasked with solving the path explosion problem and to eliminate all control flow blocks that do not influence overall analysis goals. The final stages are responsible with instruction level analysis which binds each path to a constant equation verifying the arguments passed to the system call. By means of a scalable synthetic test program generator four sets of test-benches are constructed. Each test-bench is responsible to validate its corresponding stage of the analysis pipeline. They consists of programs that are scaled on the dimensions identified through mathematical analysis of the tested algorithm and problem.

Abstract:
Static analysis of portable executable (PE) files represents a powerful tool for malware detection without running the code. A feature that can be extracted from the files during static analysis is the Shannon entropy, which is an indicator for packed binaries. Packers can be used for legitimate reasons, such as the protection of the source code, but they are often used by malware in order to hide malicious code. Multiple security solutions perform in-the-cloud file scanning in order to detect malware. However, users may be reluctant to upload their files to the cloud, especially in the case of executables that can contain private data, such as embedded resources or self-extracting archives. Homomorphic encryption (HE) represents a possible solution, as it allows the cloud servers to scan files sent by the clients directly in encrypted form. In this paper, we propose a method for computing the Shannon entropy of a plaintext string given only the corresponding homomorphic encrypted string. The encrypted entropy value can be further used, along with other encrypted static features from the PE file, by a privacy-preserving machine learning (ML) model for malware detection. In the first set of experiments, we analyzed the efficiency of the proposed method and we obtained reasonable entropy computation times for strings containing 1000 hexadecimal characters: 55.65 seconds with single-threading, respectively 19.75 seconds with multithreading. In the second set of experiments, we demonstrated that, although the Shannon entropy is generally computed at the byte level, evaluating the entropy for hexadecimal characters represents a reliable method for identifying packed PE files.

Abstract:
The mental health of social media users has started more and more to be put at risk by harmful, hateful, and offensive content. In this paper, we propose STOPHC, a harmful content detection and mitigation architecture for social media platforms. Our aim with STOPHC is to create more secure online environments. Our solution contains two modules, one that employs deep neural network architecture for harmful content detection, and one that uses a network immunization algorithm to block toxic nodes and stop the spread of harmful content. The efficacy of our solution is demonstrated by experiments conducted on two real-world datasets.

Abstract:
Malware's rapid evolution and increasing sophistication challenge traditional static and dynamic detection methods, which struggle with the high volume of new threats. Machine learning (ML) techniques offer improved detection by automatically identifying malicious patterns, but their efficacy degrades over time as malware evolves. This study addresses performance deterioration in ML-based malware detection models with time-evolving data. We propose combining feature engineering and adversarial training to enhance model resilience. By generating training data mutations, our method helps models better recognize true malware indicators, maintaining high performance over time.

Abstract:
Phishing detection is a well-studied problem, yet no universal solution, applicable in all scenarios have been found. The current research tries to leverage recent breakthroughs in the Machine Learning field, by extracting advanced features from suspicious emails. The classical features, extracted algorithmically from the email headers and body are augmented with expert features extracted by Large Language Models that are fed the email text. All these features are used in multiple Machine Learning classifiers, obtaining Accuracies and AUC scores above 90%

Abstract:
Self-Sovereign Identity (SSI) systems aim to enable citizens to control their personal data, allowing them to manage and share their identity information securely and privately. They have promising features compared to the nowadays centralized identity management, however significant challenges need to be addressed in terms of scalability, security and trust. In this paper, we propose a blockchain-based solution to manage personal identity in a decentralized, secure, and transparent manner. The blockchain technology ensures security and integrity of data, eliminating risks associated with centralized systems while the developed smart contracts enable process automation by automatically executing predefined actions related to data sharing approval when certain conditions are met, thus reducing the need for intermediaries. The system features an automated process of minting Non-Fungible Tokens (NFTs) and registering personal documents with metadata and images using the InterPlanetary File System (IPFS). The application has been evaluated for document registration and data sharing approval, showing good results in terms of transaction cost and throughput.

Abstract:
The machine learning as a service (MLaaS) concept has become increasingly popular during the past decade, as cloud platforms offer improved availability and high-performance computational resources. Nevertheless, the benefits of MLaaS come at the cost of security risks. Multi-key homomorphic encryption (MKHE) represents a solution, as it allows computations over data encrypted by multiple users under different keys. In our previous work [1], we proposed a method based on MKHE and lookup tables that allows privacy-preserving evaluations of encrypted random forest models over encrypted client queries. In this paper, we enhance the previous method, that performs only binary classification, to also allow multiclass classification. Furthermore, we demonstrate that our method can be easily adapted for the evaluation of other machine learning (ML) models and, more specifically, we provide experimental results for the Random k-Nearest Neighbors (RkNN) model. We compare the encrypted RkNN model with a traditional plaintext kNN model in terms of time efficiency and ML performance metrics. During the experiments performed on three benchmark datasets using the proposed method, we obtained classification times of the order of seconds. A minimal degradation or even improvements for the ML performance metrics were obtained in the considered scenarios.

Abstract:
We propose Area Disagreement, a novel uncertainty estimation method for Active Learning in the context of Semantic Segmentation, which capitalizes on the unique characteristics of the task. It relies on the assumption that slight alterations to the learner model’s parameters should not produce significant differences in the output, specifically in terms of predicted shapes and objects given the same input image. While a small amount of contradiction is natural, larger inconsistencies suggest that the model’s internal representation of the world is perplexed by those images, thus making them valuable for further training. Our uncertainty estimation method prioritizes images with a smaller Dice Coefficient relative to the average prediction, based on multiple Monte-Carlo Dropout inferences. Utilizing this approach, we outperformed baseline methods by a wide margin on the Cityscapes dataset, achieving 95% of the full-scale training performance using only 36% of the dataset and 97.5% of the full-scale training performance using 47% of the dataset.

Abstract:
In the context of increasing number of public threats, the development of automated real-time detection systems, capable of identifying crime activity and potential attacks in video recordings, has become a major priority. We propose accurate and efficient deep learning solutions for analyzing surveillance camera footage and signaling situations of imminent threat, based on detecting specific elements, such as theft masks and various types of weapons. We showcase the enhancement of YOLOv8 model with an attention mechanism, based on Squeeze-and-Excitation (SE), and with deformable convolution layers. The model is trained on custom datasets with optimization algorithms like Adam, AdamW and SGD. The performance tests are employed on a separate dataset, analyzing various specific metrics. Furthermore, a comparative analysis with alternative deep learning models, such as RetinaNet and Faster R-CNN, highlights the advantages and disadvantages of each model, providing insights into their performance under various conditions. Additionally, a voting mechanism is implemented to manage discrepancies in detections from multiple models. Extensive experimental results demonstrate significant improvements in detection precision and efficiency, making the proposed enhancements a viable approach for more accurate threat detection in real-time surveillance applications, while demonstrating our model’s superiority in balancing detection speed and precision

Abstract:
Flood disasters cause significant damage and loss of life globally, demanding efficient disaster response mechanisms. Visual Question Answering (VQA) systems, which integrate visual and textual data to answer user queries, offer a promising solution for flood disaster management. Previous deep learning methodologies showcased limitations in answering counting related questions. To address this, we propose DiViNeCaps (DistilBERT Vision Network with CapsuleNet classifier), a novel VQA network leveraging DistilBERT as the language model and Vision Mamba for image embedding. Our approach features an innovative attention-based fusion module, combining selfattention, gated attention, and cross-attention, followed by a Capsule network classifier. Our network improves counting accuracy and overall performance compared to state-of-the-art methods. Experimental results demonstrate that our architecture improves accuracy by 2.95% for Simple Counting and 3.38% for Complex Counting questions on the FloodNet dataset

Abstract:
Deep learning advancements have significantly enhanced computer vision applications in precision agriculture. While RGB cameras operating in visible light are affordable, they provide limited information compared to multispectral equipment. This research analyses methods to reduce the need for manual annotation when training a model using only RGB images, without compromising the model’s accuracy. We propose a semi-supervised approach where a teacher model, trained on multispectral images, generates artificial ground truth data to train a student model that operates solely on RGB images. This strategy has enabled us to achieve nearly a tenfold reduction in the required training data while maintaining similar performance metrics. Additionally, we explore the potential of segmentation foundation models to simplify the manual annotation process, reducing the need for full segmentation masks to just bounding boxes. Our findings also indicate that using multispectral images as input for the Segment Anything Model is more effective than using RGB images.

Abstract:
Online reviews greatly influence consumer purchase decisions, leading to fraudulent behavior as businesses try to improve their online presence or undermine competitors. Transformer models like BERT have become state-of-the-art in detecting fake reviews (opinion spam). This paper compares two multi-head Transformer models utilizing DistilBERT to enhance opinion spam detection: one with a single scale and the other with multi-scale granularity. Our multi-scale model, which includes an additional level of evaluation, achieved an F1 score of 0.9212 and an accuracy of 0.8663 on the YelpZip dataset. These results represent 6% and 11% improvements over the single-scale model and indicate a promising approach for improving text evaluation in opinion spam detection.

Abstract:
The exponential growth of medical data presents significant challenges in the handling, processing, and secure storage of patient records. This paper introduces an advanced digital system designed to extract and store anonymized patient medical records within a national cancer research network. Utilizing modern web technologies and advanced Natural Language Processing techniques, the proposed solution effectively manages the complexities of unstructured and semistructured data formats. The system, featuring a user-friendly web application, enhances data processing and accessibility. Validated through the successful extraction and structuring of data from over 40 patient records, it demonstrates significant potential in clinical research by reducing manual workload and improving data management. This paper details the system's architecture, functionalities, and benefits, emphasizing its impact on modernizing healthcare data management and its future applicability in the broader healthcare industry.

Abstract:
The automatic generation of poems has been a persistent challenge in Natural Language Processing (NLP), Computational Linguistics, and Artificial Intelligence (AI) since the 1960s. Generating poetry involves understanding and creating various linguistic representations such as phonetics, syntax, semantics, meter, rhyme, and literary devices to produce structured and creative texts. This complexity makes it an intriguing area for AI research, prompting continuous innovation. Despite advancements in neural network-based language models, especially Large Language Models (LLMs), automatic poem generation remains difficult. This difficulty is attributed to the training data predominantly comprising prose, resulting in models that favor prose-like outputs. Consequently, specialized techniques are necessary to develop models that can consistently produce high-quality poems. This paper presents a neural network-based model designed specifically for automatic poem generation. The proposed model outperforms other language models in generating well-defined poems with metrical verses, demonstrating its enhanced capability and consistency in poem generation.

Abstract:
This paper presents a comprehensive study on finetuning the T5-small model across multiple natural language processing (NLP) tasks using diverse datasets, including SNLI (Stanford Natural Language Inference), MRPC (Microsoft Research Paraphrase Corpus), SST2 (Stanford Sentiment Treebank), XSUM (Extreme Summarization), and IMPLI (Idiomatic and Metaphoric Paired Language Inference). Our experiments focus on optimizing model performance through various finetuning strategies and parameter adjustments. Initially, we explored the use of multiple prefixes in comparison to the introduction of special tokens to determine which strategy enhances input flexibility more effectively, facilitating user-specific customization. Then, we fine-tuned the model on all datasets except MRPC, followed by a subsequent fine-tuning exclusively on MRPC to investigate the phenomenon of catastrophic forgetting. To evaluate the model’s understanding of figurative language using IMPLI, we conducted specific tests and subsequently fine-tuned the model to improve its reliability in handling figurative language. Our evaluation encompassed various generation techniques to assess the model’s behavior post-fine-tuning. We employed distinct metrics for each task: accuracy for SNLI and SST2, accuracy and F1 for MRPC, ROUGE for XSUM, and accuracy for IMPLI. The results demonstrate the impact of different fine-tuning methodologies on the model’s performance, providing insights into effective strategies for multi-task learning with T5-small.

Abstract:
Transferring learned controllers from simulation to real-world domains is rarely observed for deep reinforcement learning applications, even physics-based tasks like the inverted pendulum being particularly difficult when transitioning to the physical world. Model-free controllers based on deep reinforcement learning have been a successful choice for many tasks. They are preferred over model-based ones due to their adaptive capability to situations not included in the initial design and for requiring less advanced expertise in control. This paper aims to apply learning-based model-free control to a real cart-pole system using a deep reinforcement learning policy transferred from simulation to a real environment without any fine-tuning involved with the dynamics model. In a simulation, the baseline deep Q-learning, an off-policy model-free algorithm, delivered an inferior performance on the Cart-Pole environment compared to the soft Q-learning, keeping the same hyper-parameters and ensuring reproducible outcomes. In the real experimental setup, besides a PD controller feasible for under-actuated systems, the soft version policy was transferred to accelerate the design process and enhance the control. Additionally, limitations related to the brittle robustness of PD controllers and the absence of accurate models in the learning phase are studied and addressed

Abstract:
Deep learning has significantly advanced object detection, with transformer-based models achieving notable improvements in both accuracy and efficiency. Despite their success, these models are inherently complex, making their decision making processes difficult to interpret. This lack of transparency poses challenges, particularly in critical applications where interpretability is essential. Although various Explainable AI (XAI) methods have been developed to address these challenges, many of them are not always straightforward for users to implement and interpret. This can make it more difficult for users to fully benefit from these methods. To improve accessibility and encourage wider adoption, this paper introduces XAI-DETR, an Explainable AI tool designed for seamless integration with transformer-based models used in object detection. Additionally, a novel evaluation method is presented to assess the faithfulness of the explanations provided by the tool, offering users greater confidence in the reliability and utility of the generated insights. The code is available at https://github.com/SuciuCaesar12/xaidetr.

Abstract:
In this article, we perform continual reinforcement learning (CRL) experiments using two benchmarks that rely on a physically simulated robot with a differential drive. CRL is a flavor of reinforcement learning (RL) in which environments themselves are taken to be non-stationary, which we model here by abrupt changes in observations and rewards, termed tasks. In particular, we demonstrate that deep Q-learning (DQN) with experience replay suffers from conflicting objectives in such a scenario: on the one hand, the replay buffers must be large to prevent catastrophic forgetting (CF), but, on the other hand, they should be small to enable rapid learning of new statistics. Taking inspiration from the domain of supervised continual learning (CL), we generalize common learning and forgetting measures in CL and show that they are highly indicative of typical effects in CRL: forgetting, intransigence, and, surprisingly, backwards transfer, i.e., the improvement of policies for past tasks by current learning

Abstract:
One of the biggest issues when training a new model is represented by wrongly annotated datasets which can lead to significant epistemic uncertainty. Such uncertainty can affect the performance and reliability of AI models, especially in medical applications where correct predictions by the model are crucial. We propose an algorithm that measures the uncertainty of each class for each image in a multi-label classification. The algorithm uses the Monte-Carlo (MC) Dropout [1] layer during the model’s training and it computes a value named ϵ-certainty per image per class. By detecting and addressing the wrongly annotated images, dataset refinement becomes possible. Additionally, the calculated ϵ-certainties are used as soft labels in model retraining. The new model is more robust to data uncertainty, thus resulting in a self-healed model from uncertain data.

Abstract:
This paper describes the research and development efforts that went into the design of a software solution capable of identifying and classifying logic gates and their connections from a circuit diagram drawn by hand. The system is based on a neural network model (YOLO) used for classification and relies on a series of steps to clean up the data. It is also capable of accurately detecting the hand drawn connections between the classified logic gates and creating an in-memory model of the drawn circuit, which it then uses to simulate its functioning. This could be very helpful in teaching activities (for a start) and maybe even in rapid circuit prototyping, if developed further. Throughout this paper, we present a few similar circuit classification works that we identified in the literature, describe the methodology employed in the classification and identification of the circuit elements and provide some details about the tests undertaken to evaluate the system’s performance and their results.

Abstract:
Designing high-performance analog circuits involves navigating intricate interactions between layout details and circuit performance metrics. Traditional methods often rely on extensive post-layout simulations to accurately predict circuit behavior. However, these simulations are computationally intensive and may become impractical for optimizing large-scale designs. To address these challenges, this paper introduces a pioneering approach that leverages convolutional neural networks (CNNs) to predict analog circuit performance directly from layout data. By analyzing layout features and parasitic elements, our CNN architecture enhances prediction accuracy while reducing the need for exhaustive simulations. This method not only streamlines the design process but also offers scalability, with promisingly significant advancements in analog layout automation and circuit performance prediction.

Abstract:
This paper addresses the challenges of implementing markerless Augmented Reality (AR) in complex manufacturing settings. Making AR systems more intuitive, robust, and adaptable is a required step to make their adoption possible in the industry. Among the hard constraints encountered in uncontrolled, real-world environments, we notably face the dynamic nature of production lines and the evolving appearance of the objects during the assembly process. Emerging deep learning (DL) methods enable 6D object pose estimation for AR registration of moving objects. However, they need a significant amount of 6D object pose ground truth data. In real-world scenarios, such a requirement cannot be fulfilled, because of two factors: the complexity of establishing an accurate 6D pose labeling procedure for large objects in a real production line and the wide variety of object states and appearances encountered along the assembly line. For this reason, it is necessary to develop alternative 6D pose estimation techniques capable of handling unseen objects. To this end, this paper introduces a novel pipeline relying on HoloLens 2 for data capture, Neural Radiance Fields (NeRF) for 3D model generation, and MegaPose for 6D pose estimation. The proposed approach enables 6D pose estimation without object-specific training or laborious pose labeling.

Abstract:
The decomposition of conceptual hierarchies is a significant task in knowledge engineering and software engineering, allowing the extraction of hidden or non-obvious knowledge pieces, or software subsystems, and the relations between them. During the last decades, many different approaches to the analysis and decomposition of conceptual hierarchies were developed, however, most of them do not pay enough attention to the semantic consistency of the decomposition, considering the internal semantic dependencies between the class attributes. Consequently, extracted knowledge items and software subsystems, can be semantically inconsistent due to the contradiction with such dependencies and model unrealistic or impossible entities or subsystems within a particular domain. Therefore, we propose a new approach to conceptual hierarchy decomposition based on the internal semantic dependencies between the properties and methods of the hierarchy members. Implementing the approach, an appropriate algorithm was developed for the decomposition of class hierarchies within the object-oriented dynamic networks, that constructs only semantically consistent subclasses and subhierarchies. To demonstrate the main scenarios of algorithm application, we provided a comprehensive example of the decomposition of conceptual hierarchy, consisting of homogeneous classes of objects.

Abstract:
Lung cancer detection remains a challenging task due to the complex nature of medical imaging and the shortage of specialized radiologists. This paper addresses the challenges of early detection of lung cancer using advanced machine learning techniques for CT image analysis. Given the limitations of current semantic segmentation methods and the lack of interpretability, we propose an innovative approach combining classification techniques with explanatory methods to enhance both precision and transparency. We developed a novel data preprocessing pipeline, a multi-stage CNN classifier training model, and a segmentation technique using concatenated Grad-CAM heatmaps. Our evaluations demonstrate significant improvements in tumour detection and model explainability, providing more effective support for radiologists. These advancements are reflected in increased segmentation accuracy, as evidenced by comparative analyses and qualitative assessments.

Abstract:
Cervical cancer remains a significant global health challenge. Precise segmentation of vascular anatomy in 3D Power Doppler ultrasound (3D--PDU) images is crucial for accurate diagnosis, effective treatment planning and continuous monitoring of the disease. Recent advances in deep learning have significantly enhanced the precision of segmentations for assessing tumor aggressiveness in various types of tumors. However, these advancements are still limited and face challenges when applied to cervical cancer vascular anatomy. Among deep learning models, U-Net excels in pixel-level image analysis, allowing for detailed mapping of vascular structures, which is vital for improving the management and outcomes of cervical cancer patients. This study proposes a framework with two stages: first, extracting the ground-truth masks of the vascular regions from cervical cancer 3D-PDU images, and second, evaluating the performance of three deep learning models in segmenting the vascular regions within these images. After the first stage is validated by gynecological clinicians, three models are evaluated: a U-Net, a U-Net variant with VGG16 backbone, and a U-Net variant with DenseNet121 backbone. The U-Net variant with DenseNet121 backbone achieved the best results on the test dataset, with a pixel accuracy of 97.08%, an Intersection over Union (IoU) of 90.63%, and a Dice coefficient of 95.08%.

Abstract:
Accurate diagnosis of brain tumors, such as glioma, meningioma, and pituitary tumors, from MRI scans is critical for effective treatment planning. This paper presents an innovative approach for brain tumor classification using hybrid models based on feature-level fusion of multiple Convolutional Neural Network (CNN) architectures, including ResNet, EfficientNet, AlexNet, VGG-Net and DenseNet. Additionally, an original CNN model was developed having a straightforward, yet highly efficient architecture. The proposed models were trained and tested on a well-balanced dataset of 5,721 MRI samples. The experimental results demonstrate that fusionbased hybrid models, integrating ResNet50 with other architectures, achieve high classification accuracy, outperforming single-model approaches. Furthermore, we implemented two majority voting schemes, which further enhance prediction reliability. This work contributes to the advancement of automated brain tumor diagnosis, offering promising solutions with potential clinical applications

Abstract:
Age-related macular degeneration (AMD), a common eye condition and a leading cause of vision loss among people age 50 and older, is the focus of this research work. By means of classical image processing techniques and deep learning methods, the identification of the neovascular membrane areas in an Optical Coherence Tomography Angiography (OCTA) is made. The identified areas are relevant for the presence of AMD allowing for the focus on biomarker extraction. The correct extraction of relevant biomarkers that can show the evolution of the disease and the effect of intravitreal injections. A dataset of OCTA images with patients suffering from AMD was created for the research work in this paper