Abstract:
Imbalanced datasets in medical imaging are characterized by skewed class proportions and scarcity of abnormal cases. When trained using such data, models tend to assign higher probabilities to normal cases, leading to biased performance. Common oversampling techniques such as SMOTE rely on local information and can introduce marginalization issues. This paper investigates the potential of using Mixup augmentation that combines two training examples along with their corresponding labels to generate new data points as a generic vicinal distribution. To this end, we propose STEM, which combines SMOTEENN and Mixup at the instance level. This integration enables us to effectively leverage the entire distribution of minority classes, thereby mitigating both between-class and within-class imbalances. We focus on the breast cancer problem, where imbalanced datasets are prevalent. The results demonstrate the effectiveness of STEM, which achieves AUC values of 0.96 and 0.99 in the Digital Database for Screening Mammography and Wisconsin Breast Cancer (Diagnostics) datasets, respectively. Moreover, this method shows promising potential when applied with an ensemble of machine learning (ML) classifiers.

Abstract:
Deep learning models are widely used for automated driving, but achieving the goal of performance in embedded Systems-on-Chip (SoCs) has a lot of challenges due to the tradeoff between accuracy and run-time. This paper addresses the task of finding a suitable neural architecture for heterogeneous Systems-on-Chip (SoCs) in order to strike a balance between accuracy and run-time efficiency. Our approach may be used to predict the inference run-time of DNN models, which is useful not only for neural architecture search, but also for the optimization of deep learning models. We propose a method for predicting the run-time of deep learning models using Graph Neural Networks, based on the Graph Convolution Network. We transform the DNN models into an acyclic graph structure, which incorporates encoded features for all nodes. Each node represents an operation of the model and all nodes are linked together. The experimental results show that our approach achieves accurate run-time inference predictions, with an overall error between 2\% and 4\% for ARM CPUs.

Abstract:
In this paper, we propose a possible framework for modeling and predicting the deformation behavior of wind turbine blades. Blade reliability is vital in the operation and maintenance of a wind turbine. A possible way to intelligently monitor the condition of the blade is to acquire and process data on the current strains inside the blade structure. In this context, the first aim of the work was to numerically simulate the equivalent strains inside the composite material of a blade. A typical rotor of a 1.5 - 3.0 MW wind turbine was considered. The rotor was simulated under various boundary conditions such as wind speed and rotation speed. The purpose of the numerical simulation of the wind turbine blade was to determine and analyze the values of the equivalent strains depending on the wind speed in the range of 6 - 20 m/s. Then, we describe the proposed framework for modeling the deformation behavior. We assume that the deformations of interest obtained from the numerical simulations are used as training samples. Our approach incorporates simulated (synthetic) data into a statistical predictive model. These data further serve to create an artificial neural network model that would be able to predict the specific deformation values at nodes of interest as a function of, but not limited to, wind speed. Experimental results are presented and discussed.

Abstract:
Minority game theory has, traditionally, been used to simulate player behavior under various conditions in a number of stock markets. However, digital markets, such as cryptocurrencies, have been largely ignored by game theory models. Using a comparative approach, with data both from traditional equities markets and from Bitcoin this article presents a model of a dollar game and compares its outcome to real-life data. The paper aims to prove that game theory can be used to predict cryptocurrency markets similarly to how it is used to predict traditional stock markets. By using historical data from the London Stock Exchange and Bitcoin the paper demonstrates that a custom implementation of a dollar game can be used to predict general market trends and the overall impact of short-term investments in Bitcoin.

Abstract:
This paper presents a method for automated vehicle trajectory planning using quantum computing. The use of quantum computing for the trajectory planning problem was studied because the current classical methods for solving the problem are computationally expensive. Simulation tests with the presented method were performed in more complex test cases than what has previously been tested with this method. Simulation tests show, that the method scales also for more complex problems. The method presented here can theoretically outperform classical methods in complexity in certain conditions, although it is not guaranteed in every condition.

Abstract:
Predicting emergency braking events in vehicles plays a critical role in enhancing road safety, particularly in the context of speeding-related accidents. In this research, we propose a convolutional neural network (CNN) approach to predict current frame emergency braking using a sequence of 20 images and 20 corresponding velocity data points as input. Leveraging the spatial and temporal information captured by CNNs, our model aims to accurately anticipate the need for immediate braking actions. We conduct experiments using the Honda Deep Drive dataset, which contains diverse road traffic scenes captured at different moments of the day and under various weather conditions. Our results demonstrate the effectiveness of the proposed approach, achieving high prediction accuracy and providing real-time warnings for potential emergency braking situations. The developed model contributes to the field of autonomous vehicles, where ensuring safe and efficient navigation is of paramount importance. By improving the prediction capabilities for emergency braking, we contribute to enhancing the overall road safety and provide valuable insights for the development of intelligent systems in autonomous vehicles.

Abstract:
Optical flow plays an important role in many applications, such as autonomous driving, where it facilitates the perception and understanding of the surrounding environment. Since the rise of deep learning approaches for optical flow estimation, much effort has been expended on new architectures and on increasing the accuracy of predicted flow. However, topics such as including semantic knowledge in model training or estimating the uncertainty for the predicted optical flow have received comparatively little attention so far. In this paper we investigate the effect of adding semantic supervision for a state of the art supervised optical flow model [1], as well as the necessary adaptations needed for obtaining a measure of flow uncertainty. In addition, our method is suitable for deployment and integration on various embedded hardware platforms for automated driving. In our experiments, we find that semantic guidance is able to improve the model performance on the KITTI benchmark by 9% on foreground objects, after having been trained only on synthetic data from the target domain. We also obtain improved flow uncertainty estimates over commonly used approaches such as model ensembles and Monte Carlo Dropout.

Abstract:
Our previously proposed linear approach for reducing the global drift of a video-based frame-to-frame trajectory estimation method corrects it at selected points in time based on the alignment of one past and the current 3D LiDAR measurements (see [7]). In this paper we improve on that method essentially by adding multi-past frame LiDAR point cloud alignment constraints and by performing the correction more often. While this significantly improves the accuracy of the estimation, it has the downside effect of decreasing the overall runtime performance. We thus study the trade-off between accuracy and speed and propose a couple of higher accuracy but real time correction versions. Their evaluation on the KITTI dataset results in an overall performance falling into the so called SLAM-nonSLAM gap.

Abstract:
The paper addresses the challenge of enabling secure and privacy-preserving data aggregation within decentralized business communities. The specific application scenario involves small communities of big companies that seek to access detailed aggregated data from their community without revealing their individual contributions to competitors or any centralized authority. The absence of a neutral and trusted organization to collect and aggregate data necessitates the development of a decentralized solution. To meet these requirements, we propose a novel P2P algorithm that ensures efficient connection establishment among nodes while maintaining privacy and anonymity. By leveraging the network topology and employing specific connection request types, nodes can establish connections without disclosing their specific business details. Experimental results validate the algorithm's effectiveness of securely constructing a cyclic overlay network. The solution offers a robust and scalable approach to data aggregation, enabling actors to access aggregated information without compromising their individual data or the confidentiality of their contributions.

Abstract:
Anomaly detection is the technique of finding outof-the-ordinary occurrences inside datasets. A multitude of anomaly detection methods was developed for this purpose in a wide variety of domains (e.g., economy, industrial processes, health, environmental monitoring, etc.), using as support more generic artificial intelligence and signal processing algorithms. The performance of these algorithms is strongly influenced by the initial dataset characteristics and by the hyperparameter settings that users choose for each algorithm. There is no apriory best solution for a given domain or type of dataset; therefore, the researchers must test a multitude of methods with multiple hyperparameter settings. The research community does not yet have a multi-perspective assessment of the performance of these algorithms. Most of the work is focused on very specific problems or domains, and the algorithms are evaluated only from one or a few perspectives, using the ”default” settings, without hyperparameter optimization. In this article, we will discuss the findings obtained from an extensive anomaly detection benchmarking that covers 34 anomaly detection algorithms, tested on several relevant datasets from a wide area of domains. The results are presented from multiple perspectives. For each algorithm and dataset, a hyperparameter optimization approach is used in order to establish the optimal settings for each algorithm in each test case. This study discloses both the merits and limitations of the various techniques and proposes a taxonomy of anomaly detection algorithms based on the nature of the problem and the characteristics of the input data.

Abstract:
The phenomenon of Gravitational Wave (GW) analysis has grown in popularity as technology has advanced and the process of observing gravitational waves has become more precise. Although the sensitivity and the frequency of observation of GW signals are constantly improving, the possibility of noise in the collected GW data remains. In this paper, we propose two new Machine and Deep learning ensemble approaches (i.e., ShallowWaves and DeepWaves Ensembles) for detecting different types of noise and patterns in datasets from GW observatories. Our research also investigates various Machine and Deep Learning techniques for multi-class classification and provides a comprehensive benchmark, emphasizing the best results in terms of three commonly used performance metrics (i.e., accuracy, precision, and recall). We train and test our models on a dataset consisting of annotated time series from real-world data collected by the Advanced Laser Interferometer GW Observatory (LIGO). We empirically show that the best overall accuracy is obtained by the proposed DeepWaves Ensemble, followed close by the ShallowWaves Ensemble.

Abstract:
Machine learning has triggered a paradigm shift, facilitating the gathering of relevant information from a large amount of seemingly diverse data. The primary goal of this research paper is to emphasize the effectiveness of a varied set of data analysis methods used to extract meaningful intelligence from a complex data set gathered by experimenting with household appliances in a controlled environment. The methods differ depending on the goal of the analysis and the characteristics of the data: machine learning, statistical modeling, and natural language processing. In this paper, we demonstrate how visual representations of the information can be a deciding factor when analyzing experimental data. The impact of computations on the semantic and structural form of the data, as a pre-processing step, is also detailed.

Abstract:
The Lightweight Cryptographic Key Distribution Protocol (LOKI) was initially proposed as a solution that enables computationally restricted automotive control units to periodically update short-term authentication keys. After analysing the protocol messaging from a security and a formal point of view, we found that the protocol lacks critical security properties. The work at hand introduces an improved protocol version that intents to address the initial design flaws. Moreover, a formal automated verification was conducted on both protocols, first to demonstrate the shortcomings of LOKI, and secondly, the correctness of the improved variant. Finally, the improved protocol is formally verified with a Burrows–Abadi–Needham (BAN) logic analysis.

Abstract:
Enhancements in driving equipment have been instrumental in promoting autonomy and flexibility. However, the act of driving at night poses challenges to individuals due to personal and environmental factors. Particularly when navigating a curve in a car at night, the fixed illumination angle of the vehicle's lights may fail to adequately illuminate the blind spot in the turn. To address this limitation and enhance safety in situations with limited viewing angles, this research paper introduces a calculation for an adaptive front-lighting system (AFS) for automobiles, controlled by an electric control unit. This paper conducts a comprehensive investigation into the operational principles that govern the adaptive front-lighting system (AFS) implemented in automobiles, with a particular focus on its performance during curve negotiation. It formulates a precise mathematical model to accurately portray the system's behavior and presents an extensive overview of its overall design. Additionally, the article delves into the intricate procedural aspects of the software utilized in this system. The AFS system adjusts the angle of the light axis based on variations in the steering wheel angle and speed, thereby augmenting the driver's vision and overall safety. By employing an adaptive driving beam consisting of multiple LED segments that can be individually activated, deactivated, or dimmed, the illumination beam can be customized to suit diverse driving scenarios, including mitigating glare when faced with oncoming traffic. A key advantage of utilizing LED matrix lights for adaptive front-lighting in vehicles is their independence from mechanical components for beam shaping. Consequently, the present study explores the optical principles required to achieve the desired angular resolutions for the matrix segments and for the adaptive front-light system. Furthermore, a specific solution is presented, involving the utilization of a 3x11 matrix arrangement of ceramic light-emitting diode (LED) packages assembled on a single board. Additionally, this paper offers comprehensive insights into the operational principles of the AFS, including its model and hardware circuits. By addressing these critical aspects, this research aims to contribute to the advancement of compact and cost-effective solutions for the realization of an efficient and accessible adaptive front-lighting system.

Abstract:
Vital guidance information for autonomous driving cars can be extracted from road markings. The main challenges come from the wide range of traffic scenarios, lighting and weather conditions. In this work, we look into the problem of road marking detection with object detection neural networks on the CeyMo Road Marking dataset. We develop an architecture named SE-YOLOv7-tiny, which is based on YOLOv7-tiny, as it is a lightweight object detector that can be deployed on edge devices. To improve model performance, we propose several adjustments to YOLOv7-tiny. Firstly, we replace the Leaky-RELU activation function with Mish. Then we fine-tune data augmentation parameters and introduce into the pipeline the mosaic data augmentation technique. The final adjustment that we propose is to introduce squeeze-andexcitation blocks for features fusion into the neck part of the architecture. We evaluate the inference speed of the obtained models on NVIDIA RTX 2070. The experimental results show that our method achieves a macro F1-score of 89.59% and scenario overall F1-score of 93.32%. Compared with previous state-of-the-art (SSD-Inception-v2) on the object detection task, our model improves macro F1-score by 6.71% and scenario overall F1-score by 8.13%, while achieving twice the FPS.

Abstract:
The 3D reconstruction of a real-world scene, usually represented as a textured mesh, supports the fusion of various information and can solve complex problems. For example, the fusion of the 3D textured mesh with a relatively small set of semantically annotated input images can generate a supplementary semantic mesh. The two meshes can be used, based on a set of consecutive camera positions, to generate novel RGB, depth, or semantic images. However, these meshes are not able to represent the dynamic objects, as these objects tend to vanish in the 3D textured mesh construction process. The main goal of this work is to provide a solution to generate training sequences including dynamic entities. Therefore, for a given camera pose and time instance, a new RGB, depth or semantically annotated image can be generated using a mesh instance that inserts the dynamic entities corresponding to the given timestamp. The proposed solution utilizes Blender, an open-source software tool, to place and animate the photorealistic mesh models of the different dynamic entities in the 3D mesh representation of the realworld scene. We use this tool to manage environment occlusions and shadows, and to mimic the properties of the scene. After the photorealistic rendering, we obtain a set of images of the dynamic objects and their shadows. We propose a technique to transfer the generated dynamic information into RGB, depth and semantic images. The method focuses on entities such as moving cars and pedestrians, but any dynamic entity is suitable. We used a subset of the UAVid dataset to test the practical viability of the solution for supervised semantic segmentation training. Experimental results show that using the enriched images, as opposed to the initial images, increases the performance on the semantic segmentation task by 10.77% mIoU.

Abstract:
The proposed Multi-Stream TCP (MS-TCP) transport protocol challenges the traditional one-to-one relationship between the Transport Control Protocol (TCP) connections and full-duplex streams. This paper explores the feasibility and advantages of allowing a single TCP connection to support multiple streams through distinct sockets, with a focus on enhancing the efficiency of concurrent web transactions. The MS-TCP protocol addresses the increasing demand for efficient utilization of TCP network connections in today's network interoperation, dominated by HTTP-based transactions, offering a legacy-aware architecture alternative to the disruptive HTTP/3 QUIC solution. By relaxing the strict connection-stream relationship, MS-TCP introduces the concept of "channel", a single TCP-based connection that can accommodate multiple concurrent streams. The paper presents a detailed examination of the design, architecture, and implementation choices involved in realizing the requirements of MS-TCP. These requirements include adherence to POSIX Standard APIs, seamless integration with existing network middleboxes, mitigation of Head-of-Line (HoL) blocking issues, enabling zero roundtrip time, and skipping the slow-start state when activating new streams.

Abstract:
The recent events in the world, such as the Covid19 pandemics raise the importance of tracking room occupancy. The importance of estimating the number of persons present is not related only to the epidemic scenarios and attempts to avoid people to people contact with the goal to stop the spreading of the diseases. This importance can be expanded to the scenarios when we want to avoid the presence of a higher number of people in certain areas to offer their safety and security for a variety of reasons, e.g. employee welfare, hazardous material presence, etc. Although there are numerous approaches to tackle this problem, this paper only deals with the usage of Wireless Sensor Networks (WSN), with standard or common nodes set to estimate the number of persons present in the targeted space.

Abstract:
CERN is an European research facility that operates the largest and most powerful particle accelerator in the world, called the Large Hadron Collider (LHC). The LHC hosts four large experiments studying proton–proton and heavy-ion collisions at high energies. One of the four experiments is ALICE, specifically designed to study the physics of heavy-ion collisions. The experiments collect hundreds of PB of data each year and require complex mechanisms to maintain and distribute these data over many different storage elements around the world on the Worldwide LHC computing Grid. This paper describes a data management system used by ALICE, which maintains the data workflow from the experimental apparatus to the primary 150 PB disk storage. The system operates in real time and fulfills stringent requirements for transfer speed, transfer control, and monitoring. It is also completely automatic and autonomous, with sophisticated fault detection along the entire data transfer and registration process.

Abstract:
As the realm of Big Data gains momentum, driven by the ever-increasing volume of information that continuously needs to be stored and processed, the problem of availability becomes pressing. The current work addresses this high-impact problem, by using a novel approach to build a decentralized, distributed database for user credentials. This implementation leverages the peer-to-peer DHTs, as their decentralized architecture enables efficient data management and replication in a distributed environment. In the context of the broad market of traditional database technologies, this work can be integrated into applications as the underlying storage solution, in a seamless and transparent manner for the application that uses it. High availability and fault tolerance, inherited from the DHT, are exhibited through experimental evaluation and rigorous analysis of the results, illustrating the viability of this original solution.

Abstract:
This paper presents an optimization and multiobjective evaluation of deep learning (DL) models to improve pallet localization with an RGB-D camera in scenarios of forklift insertion. To this end, we experimentally evaluate three distinct DL models: Detectron2 for combined detection and segmentation, YoloV5 for detection only, and a combination of YoloV5 and UNet for detection and segmentation. Through the automatic hyperparameter optimization process on a custom dataset, 30 configurations were selected, each demonstrating a unique trade-off between precision and speed. Out of these, three Pareto optimal models were chosen for a more detailed analysis, considering inference speed and localization errors along each orthogonal axis. The results suggest that the proposed hybrid model combining YoloV5 and UNet exhibited a promising balance of speed and accuracy, making it suitable for real-time applications. Finally, the proposed model is demonstrated for a fork insertion task in a new environment.

Abstract:
The ability to recognize and classify 3D object models into discrete categories is becoming more and more important. It has applications in a wide variety of fields ranging from augmented reality to robotics and self-driving cars. In this paper, we present a novel approach to 3D Object Recognition that aims to transform 3D objects into 2D representations that preserve depth as well as possible. The transformed models are then fed into Vision Transformer-based neural networks that are able to classify them into 40 discrete categories. For the training, we use Princeton's ModelNet40 data set which contains 12, 311 pre-aligned models labeled into 40 categories. We believe our approach is important as it allows us to use recent breakthroughs in the task of 2D classification, such as Vision Transformers, for the task of 3D classification.

Abstract:
The use of Augmented Reality (AR) devices like HoloLens has gained significant acceptance for training assembly line operators in various industries. When employing Computer-Aided Design (CAD) models to create assembly line instructions for training purposes, preserving all redundant information becomes unnecessary. Utilizing simplified CAD models leads to improved run-time performance of the applications in which they are employed. This paper introduces a novel solution to simplify CAD models by identifying continuous chains within the CAD model, with the help of deep learning techniques. The proposed method makes it possible to suppress such features that do not impede visualization when removed.

Abstract:
Sign language is an essential form of communication for people with hearing disabilities. However, effective communication traditionally required the use of trained translators. With recent advances in AI and virtual reality, technology seems to provide a viable means of facilitating communication with sign-language users. This paper describes a solution developed around the detection and classification of the user's hand signals and shows its integration within a system meant to facilitate communication with hearing impaired individuals. The system relies on the capabilities of a virtual reality headset to capture the user's hand signals and then employs a machine learning model in order to correctly classify them. Within this paper, we discuss some of the relevant technical aspects of the system as well as provide insight into its observed capabilities.

Abstract:
The Lightweight Cryptographic Key Distribution Protocol (LOKI) was initially proposed as a solution that enables computationally restricted automotive control units to periodically update short-term authentication keys. After analysing the protocol messaging from a security and a formal point of view, we found that the protocol lacks critical security properties. The work at hand introduces an improved protocol version that intents to address the initial design flaws. Moreover, a formal automated verification was conducted on both protocols, first to demonstrate the shortcomings of LOKI, and secondly, the correctness of the improved variant. Finally, the improved protocol is formally verified with a Burrows–Abadi–Needham (BAN) logic analysis.

Abstract:
A large number of malware samples use Domain Generation Algorithms (DGA) for locating and connecting to the Command and Control server. This technique enables the malware authors to avoid the C\&C domain being seized and makes it difficult for security solutions to blacklist all the possible generated domains. Since the malicious actors register only a small fraction of the generated domains, DGA-enabled malware will generate a large number of unresolved DNS queries. Unfortunately, security solutions cannot rely solely on this heuristic, as computing systems can also produce unresolved DNS queries for benign reasons. Checking whether an unresolved domain is generated by a DGA can be done using Machine Learning algorithms, but the model usually requires high computing power, either locally or in the cloud. We propose a low footprint Machine Learning model that is able to detect generated malicious domains without the need for a great deal of memory and overall computational power. Furthermore, our model enables resource-constrained devices such as home routers to protect themselves and the local network. The experimental results show that our model has a close performance to the state-of-the-art classifiers while using just a fraction of their resources.

Abstract:
Employing Machine Learning (ML) methodologies to address the issue of malware detection is an extensively researched subject. Researchers place particular emphasis on selecting an appropriate model and optimizing its performance through effective hyperparameter configuration. Ember is one of the primary datasets used for training and testing ML models for malware detection. In this study, we are interested in whether selecting a specific subset of the features defined in the Ember dataset would result in better-performing models. We seek to identify those feature groups that carry the most helpful information for the model to leverage and show that using only a combination of them can yield better performance than using the entire set of features.

Abstract:
Machine learning methods are commonly used for malware detection to respond to the ever-increasing number of malicious programs. Detection models are usually trained on full execution logs and are unsuitable for real-time malware detection during execution. We build a real-time malware detection system based on process behavior combined with artifacts from process memory that aims to show that real-time malware detection using machine learning is feasible. We optimize our system by separating detection into two stages, one stage based on classical machine learning models and the other on a deep neural network. We show how we train our models and how we evaluate them. We introduce a novel way to evaluate detection models that aims to improve system performance.

Abstract:
This paper presents a specialized Single Shot Multibox Detector (SSD) [9] tailored to the unique challenges of detecting low-textured objects in complex scenes. We harness the Fruit 360 dataset [10], featuring low-textured images of various fruits and vegetables, for both training and validation purposes. Our primary objective is to enable the integration of this streamlined SSD model [9] into mobile devices for the enhancement of mixed and augmented reality experiences. Remarkably, our custom model delivers a fourfold increase in processing speed compared to the original SSD [9], all while preserving or even enhancing detection accuracy on the designated dataset. Implementation is carried out in Python using TensorFlow.

Abstract:
This research paper presents the implementation and evaluation of a Total Variation (TV) layer within a deep learning framework for image denoising tasks. The TV layer is based on Chambolle's projection method and aims to enhance the performance of a baseline denoising network. The study investigates the effectiveness of incorporating a single TV denoising layer, which adds only six parameters to an arbitrary baseline autoencoder. The resulting models are evaluated across three datasets with different Gaussian noise. By comparing an arbitrary network architecture with an identical counterpart model that incorporates the TV denoising layer, the results indicate significant improvements in performance. The magnitude of these performance gains diminishes as the variance of noise increases. Notably, for lower Gaussian noise variance such as for a standard deviation of 25, the gain in terms of the mean peak signal-to-noise ratio (PSNR) metric is estimated to be approximately +0.89 dB, a discernible improvement that is visually perceptible. Although the resulting model has no significant increase in terms of model size, it does present significant increases in complexity.

Abstract:
In this work, we explore the use of a deep learning neural network architecture for doing object detection of medical fluid samples, from color images. The ability of reliable identification and localization of liquid phases inside sample containers can help in automating significant parts of the work carried out in medical laboratories and hospitals. Our focus is on tracing the liquid levels of the plasma portion of whole blood, which may be further used in aliquoting, platelet rich plasma treatments automation, complete blood count analysis and others. The majority of sample containers are transparent and may contain or may not, one or more barcodes attached on the external surface, in addition to the manufacturer's labels. Consequently, the available clearance window to view the liquid contents is of variable size and shape, and may be sufficiently large only from few view angles of the sample container. To address the clearance window challenge, we introduce a new input representation technique, named vertical image stitching, that takes multiple images of a rotating sample container, extracts a region of interest from each of them and then stitches them together vertically in a single output image, called vertically stitched image. Based on the vertical stitched image, we introduce a new data augmentation technique named vertical stitch permutation, which enables obtaining 10 times more sample variations in the dataset. Finally, we convert the 2D bounding box localization computation into 1D points localization by removing from the loss function the bounding box components corresponding to the box height. We performed experiments on a dataset of 976 unique samples, containing 17 sample containers from 11 manufacturers. We trained YoloV4- tiny with the above mentioned techniques and achieved a mean average precision of 98.09% (at intersection over union threshold of 0.90) with an average intersection over union of 95.18%.

Abstract:
Semantic segmentation is a critical task for any system that requires scene understanding, as it computes the objects' class at pixel level. We present VTSeg, a novel video semantic segmentation framework that exploits spatio-temporal information between consecutive frames to achieve a higher segmentation quality. VTSeg leverages semi-supervised learning, eliminating the need for additional training data. The architecture employs a unique adaptation of Transformers to video semantic segmentation by making an analogy between image representation spaces and natural languages. The encoder and decoder incorporate an innovative video attention module and generate multiscale features. Additionally, we introduce a novel segmentation head that uses super-resolution techniques. Compared to other video semantic segmentation approaches that include optical flow, we achieve a 3% mIoU improvement on the Virtual KITTI 2 dataset, even when ground truth is used directly for optical flow. Compared to image segmentation networks, VTSeg keeps the same 3% mIoU distance despite using 75% fewer labels. On the Cityscapes validation set, we obtain 76% mIoU using half the input resolution, without including any of the usual quality enhancing techniques such as ImageNet pre-training, test set augmentation, or sliding window inference. VTSeg framework learns powerful representations from both labeled and unlabeled data, making it a suitable video adapter able to boost the results of any image-based semantic segmentation network.

Abstract:
The proliferation of WiFi access points has significantly transformed our interactions with the digital world. However, the crowdsourced collection and dissemination of WiFi access point data have raised considerable concerns regarding privacy. The availability of such data can facilitate the identification of access points, and the presence of identifiable information, including names, can enable inferences regarding the location of an individual's residence, thereby jeopardizing the privacy of sensitive personal information. In response to these concerns, we propose a novel approach to privacy-preserving crowdsourced WiFi data publishing, utilizing differential privacy. This approach ensures that the released data maintains the confidentiality of WiFi access point owners while enabling useful analysis of the network data. The objective of this article is to investigate the challenges and possibilities associated with the use of differential privacy in crowdsourced RSS data publishing for WiFi access points, with an emphasis on preserving privacy. The proposed method is evaluated in a simulated environment using MininetWiFi.

Abstract:
Whenever one receives a phone call, the Caller ID is the first bit of information that is being processed. Based on this information, one may choose to accept the incoming call or to reject it. Furthermore, the Caller ID is the element that sets the trust premise of the call, even before the actual conversation begins. Unfortunately, the value of the Caller ID is neither guaranteed in any way by the telecommunications company nor is it imposed and can easily be spoofed by a malicious actor in order to start an attack that wouldn't be possible otherwise. This paper presents a method of identifying a spoofed Caller ID, method that is applicable to smartphones with no need of additional hardware, no additional running costs or the implication of third-party entities.

Abstract:
This paper addresses the problem of data privacy in distributed applications using machine learning models in the context of modern privacy laws, such as GDPR, by proposing a machine learning model hybridized to use homomorphic operators to perform model evaluations on predicted data. The paper presents the general approach of building such a model using frameworks such as TesSeal, Keras, Microsoft SEAL, and their integration in a web-distributed application. Due to the complexity of the homomorphic operations, only the prediction is proposed to be executed on encrypted data. An experimental platform is developed where a machine learning algorithm is trained to predict diabetes based on laboratory tests. The proposed models that use homomorphic operation to run on encrypted data are evaluated and compared to the same models that operate on plain-text data to determine the cost of encrypting the data from the point of view of model accuracy and execution time. Experiments show that both the model that runs on encrypted data and the model that runs on plain data have similar accuracy. However, the execution time of the model that runs on encrypted data is 10 to 20 times higher than of the model using plain text data.

Abstract:
In cybersecurity, understanding an attacker's tactics and techniques gives the defender an edge. MITRE ATT\&CK provides a framework that describes 14 tactics commonly used by attackers in the act of breaching an organisation infrastructure. Splunk, as a log management and analysis platform provides a comprehensive query language that can be used to trigger security incidents. This paper attempts to solve the problem of annotating Splunk queries with MITRE tactics, bridging the semantic gap and providing more insights to security analysts and Security Operations Center operators. The proposed solution uses Machine Learning classifiers to solve this problem and was tested on a dataset containing more than 5000 Splunk queries.

Abstract:
An imbalanced ratio of body fat to muscle mass increases the risk of fatal illnesses, such as cardiovascular diseases, diabetes, and cancers. The analysis of this ratio requires the detection and segmentation of fat and muscle in medical images and precise segmentation of body composition helps to analyze more efficiently. Many existing commercial segmentation methods are either manual or semi-automatic, requiring intervention from reviewers or specialists. This paper introduces a fully automatic machine learning method that utilizes a 3D U-Net based Convolutional Neural Network to automatically segment muscles, adipose tissues and bones. To the best of our knowledge, this is the first application of a fully convolutional neural network(CNN) architecture for whole-body segmentation using magnetic resonance imaging (MRI). The 3D whole-body MRI scans were obtained using a 3D two-point Dixon VIBE sequence, which is the latest advancements for obtaining four types of MR images. The acquired data from 14 subjects fed into the U-Net model, which outputs the volumes of muscles, bone, and adipose tissues. The performance of the designed network was quantitatively evaluated using the Dice coefficient (Dice). The experimental results demonstrate that the proposed method is both simple and robust when compared to other studies on human. The average Dice on selected data reached 0.86.

Abstract:
Kidney tumor grade identification by means of feature based classification combined with semantic segmentation of Computed Tomography (CT) images targeting tumor region extraction are the two main contributions of this paper. For all the patients involved in the study three phases of the CT examinations are considered: native, arterial and venous phase. Medical specialists have annotated the images marking the regions in which tumors reside for each of the three phases. On each type of phase deep learning based segmentation has been applied in order to identify the tumor area. The differentiation among various tumor grades by means of deep learning is extremely challenging due to the small amount of available image data and due to the high similarity between different tumor grades. Hence feature based classification is applied to segmented regions in order to distinguish among the four tumor grades. The accuracy of the recognition varies from 60% up to 90% depending on the tumor grade, with first grade tumors being more difficult to be recognized while fourth grade tumors are correctly identified for most patients.

Abstract:
Age-Related Macular Degeneration (AMD) represents one of the leading causes of blindness for individuals over 65. Optical Coherence Tomography represents a noninvasive examination that not only can help clinicians diagnose multiple retinal abnormalities, AMD included, but also monitor the progression of the disease. There are still unmet needs in terms of personalized treatment for patients suffering from AMD, being a disease that displays individual diversity in terms of its progression and outcomes. We propose a method that will use deep learning methodologies to analyze patients' disease progression and possible outcomes only based on the OCT scans that are taken during the initial first two examinations. This paper will propose an architecture that will help medical professionals with their involvement in the administration of antiVEGF injections, which are the standard treatment for advanced neovascular AMD. The above-mentioned architecture is based on features extracted from B-scans of an OCT volume and through transfer learning practices will predict the total amount of injections required for an individual who is under treatment, the next visit decision to administer the injection, as well as visual acuity measured at the last visit.

Abstract:
Detecting polyps in endoscopic images is very important for early detection of colorectal cancer, hence the importance of proper training of practitioners. Preparing a good set of images for training them is a challenging task, due to lack of comprehensive image sets. To address this, we explored a method based on Deep Learning to automatically detect polyps in colonoscopy images and generate the corresponding masks, such a solution providing the ability of turning unlabeled data into an important learning resource. We explored different augmentations techniques and used a FCBFormer based model, which was trained using different hyperparameters, managing to obtain two main versions that are noteworthy, one that is using the AdaBelief optimizer and the weighted sum between the Binary Cross-Entropy loss and the Dice loss as the loss function, which leads to a value of 0.93 for Dice and 0.99 for Precision, outperforming the baseline FCBFormer architecture. On the other hand, employing the Adam optimizer and Binary CrossEntropy loss leads to better Recall and Precision values (0.95 and 0.998 respectively) at the expense of a slight reduction of the Dice score from 0.93 to 0.92.

Abstract:
This research focuses on Federated Learning for intent classification as a core element to any conversational system. Since the level of accuracy needed for intent classification is typically quite high and considering the challenges with accessing enough conversational data with which to train and develop a classification models, we exploit Federated Learning to evaluate its feasibility in improving classification accuracy over more traditional, centralized approaches. The research findings are positive, showing an increase in performance across all metrics and achieving a higher accuracy on average across all the tests conducted.

Abstract:
In recent decades, technological advancements in the field of Artificial Intelligence (AI) have opened up new opportunities for generating audio content based on text. In particular, Text-to-Speech (TTS) transformation models have become the most popular method for generating speech from written text. These models can be utilised in applications such as audiobooks, various virtual assistants, and video games, and their performances continue to improve, with generated results becoming increasingly close to human voice. However, there are still many challenges regarding generating as natural voice as possible and appropriate expression of emotions. In this paper we will focus on TTS synthesis, understanding the state-of-the-art and exploring possibilities of training these models for languages other than English, diverse voices, and improving the quality of voice generation. We introduce an approach for Romanian language TTS synthesis using generative deep learning models. Our approach yielded promising results on Romanian language datasets, demonstrating the effectiveness of our methodology.

Abstract:
The use of Natural Language Processing Algorithms (NLP) for automation purposes in various applications is frequently encountered recently. Some research managed to identify the dominant emotion from a text using neural networks (ANN), Random Forest (RF) and Support Vectors Machines (SVM) while other studies classified documents with the aim of mechanizing the extraction process of information. This paper presents a study on different Natural Language Processing Algorithms (NLP) used for automation of a virtual bookstore. The objective was the application of artificial intelligence algorithms to streamline the necessary processes of an online bookstore or library, taking into consideration the short descriptions and summaries of the books. The aim is to improve the results obtained by supervised algorithms using various techniques such as aggregating unsupervised classifications. In order to boost their understanding of natural language, GPT is used to enhance the dataset by adding additional context. Finally, some of the methods utilized in improving accuracy can be used to create a personalized recommendation system that suits each reader's needs.

Abstract:
The aim of this study is to evaluate the performance of AI-assisted programming in actual mobile development teams that are focused on native mobile languages like Kotlin and Swift. The extensive case study involves 16 participants and 2 technical reviewers, from a software development department designed to understand the impact of using LLMs trained for code generation in specific phases of the team, more specifically, technical onboarding and technical stack switch. The study uses technical problems dedicated to each phase and requests solutions from the participants with and without using AI-Code generators. It measures time, correctness, and 'technical integration' using ReviewerScore, a metric specific to the paper and extracted from actual industry standards, the code reviewers of merge/pull requests. The output is converted and analyzed together with feedback from the participants in an attempt to determine if using AIassisted programming tools will have an impact on getting developers onboard in a project or helping them with a smooth transition between the two native development environments of mobile development, Android and iOS. The study was performed between May and June 2023 with members of the mobile department of a software development company based in ClujNapoca, with Romanian ownership and management.

Abstract:
Neurons in the brain fire action potentials or spikes to encode information and communicate with each other. The firing can be in tonic mode, where individual spikes are emitted at relatively distant time intervals, and in burst mode, where neurons typically fire trains of spikes in short temporal succession. Functionally, neuronal bursts may be involved in reliable information transmission or specialized information encoding schemes. Understanding the dynamics of neuronal burst activity plays a crucial role in unravelling the complex mechanisms underlying neural information processing. However, detecting bursts in extracellularly recorded neural data is far from trivial. There are many methods for burst detection, but none was adopted universally, and they are based exclusively on the evaluation of spike times. This article aims to compare these burst detection algorithms against each other from several perspectives in order to establish the most suitable and robust one. Moreover, we propose a novel burst detection method considering additional features that contain as much information as possible from the recording.

Abstract:
Spike sorting is a technique in the field of neuroscience applied to identify and classify the action potentials (spikes) of neurons in extracellular recordings. This method is important considering that the simultaneous activity of several neurons can be recorded through the use of electrodes, but the recorded signals are often mixed. Extracting the characteristics of individual spikes provides valuable information about the identity of neurons that generate them, the type of cells involved (inhibitory/excitatory), encoding of information in spike rates or spike times, and so on. Therefore, unmixing the spikes recorded on an electrode via spike sorting is crucial. Here, we apply a novel method to extract features of mixed spikes, namely the Superlet transform, which enables the computation of spectral characteristics with a higher resolution. We use machine learning to determine which features of the Superlet spectrum contain the most information about the shapes of individual spikes, thereby enabling their unmixing during spike sorting.

Abstract:
Spike sorting strongly relies on the exploitation of similarity of features extracted from waveforms of superimposed spikes of neurons. Main issues of such a task are given by overlapping clusters due to imperfect features and cluster imbalance that is inherent to neuronal data, as different types of neurons exhibit various firing rates. Current techniques are still highly affected by these drawbacks. To address these issues, a Self-Organizing Map-based technique is proposed, which is a neural network capable of unsupervised learning from which the unified distance matrix is extracted and employed in an additional postprocessing step that allows the proposed method to substitute both the feature extraction and the clustering steps of the spike sorting pipeline. The performance evaluation encompasses synthetic datasets, which are publicly available for the purpose of benchmarking. These datasets contain a high number of clusters, allowing for comprehensive testing and comparison of the proposed method. The results of the evaluation indicate that the proposed method outperforms other state-of-the-art techniques in terms of spike sorting performance.

Abstract:
This paper is concerned with developing an analysis tool for electroencephalogram (EEG) data. It explores the possibility of using the self-organizing map algorithm as starting point in a symbolic analysis based approach, with the aim of reducing the challenges brought by the high dimensionality and complexity of EEG data. The solution represents a pipeline integrating different steps for accomplishing an in-depth analysis: self-organizing map training, clustering, color sequence generation, pattern specificity index, pattern triggered average and peristimulus time histogram computation.

Abstract:
From the release of BERT in 2018 to the newest large language model of OpenAI, GPT-4, the corpus size of the training data has increased to the order of trillions of tokens. However, the models' capabilities to expand and generalize their acquired knowledge remain a challenge. This research aims to analyze the ability of models to generalize relation types for the task of relation extraction, by clustering them in categories and training the model on half of the relations in a cluster. The generalization capabilities will be tested against zero-shot, oneshot, and few-shot learning, introducing relations from the other half of the cluster, to analyze how much data a model needs to achieve satisfactory results on generating the unseen relation types.

Abstract:
Recent breakthroughs in the study of Denoising Diffusion Models have constituted a driving factor in advancing the field of Computer Vision through the aid of generative Machine Learning. Challenging problems, such as image or video synthesis, become intertwined with Natural Language Processing techniques to push the boundaries of computational capabilities and foster human ingenuity. We present an approach to solving text-guided video completion, a difficult task of producing the temporal future progression of an image, conditioned by textual description. Our proposed framework applies the latest academic leaps in Text-to-Image and Text-to-Video generation to perform Image Animation guided by text, an insufficiently explored topic due to its strenuous nature. We present how adapting generalized pre-trained Text-to-Image models can achieve good performance solving the described problem in a Zero-Shot context. Our experimental evaluation shows that the proposed solution is robust, capable of animating any source image guided by a language prompt, by inheriting the generality of the backbone pre-trained Text-to-Image model.

Abstract:
On-line press is infested with poor quality content that uses misleading headlines in order to increase web traffic and gain popularity and profit. This phenomenon is called clickbait and impacts the usability and, in some cases, even the security by exposing the user to malicious advertisement. This paper proposes an automated solution for detecting clickbait by using Natural Language Processing techniques to extract features from both the article title and the article content. The extracted features are used with a set of Machine Learning classifiers which are fine-tuned to improve the performance metrics. The resulting model shows good results and can be integrated in a news aggregator that selects only quality content for the user to read.

Abstract:
Emotional speech recognition (ESR) seeks to identify human emotions by capturing voice attributes, which can be invaluable for understanding speakers' moods, attitudes, and feelings. This study investigates affective computing for anger recognition in audio signals. We aim to develop a robust and efficient model capable of detecting anger across various real-world scenarios by employing feature extraction, and virtual enlargement techniques. The experimental setup includes multiple feature sets with different computational costs, and virtual enlargement strategies to simulate different environmental conditions. Results show that models trained with augmented data outperform those trained with non-augmented data, highlighting the importance of data augmentation techniques in enhancing model performance and generalization. Furthermore, additional experiments were conducted to confirm the robustness of data augmentation techniques, which consistently demonstrates their positive impact on model performance. Finally, the study contributes to the field of ESR, particularly anger recognition, and its potential applications in violence detection and public safety. The findings demonstrate the effectiveness of feature extraction, virtual enlargement, and data augmentation techniques in improving model performance across various environmental conditions, thereby enhancing the ability to identify anger in audio signals.

Abstract:
Heart failure is a growing concern due to its high incidence nowadays, also representing a major cause of morbidity and mortality worldwide. In this paper we propose a web-based platform that incorporates both the clinical data prediction aspect and the continuous monitoring of the heart health. We implement multiple machine learning models that can support the doctors in the process of classification between a healthy and unhealthy situation. The platform benefits from an ETL (Extract, Transform, Load) sub-system that processes biometric data from smart wearables and displays it in customizable dashboards for a more illustrative visualization. The prediction service integrates three Machine Learning (ML) techniques, namely Logistic Regression, Naive Bayes Classifier, and a custom Artificial Neural Network responsible for making classifications on the monitored data. The results illustrate that the proposed solution's usage for performing remote monitoring and heart health assessment is feasible, obtaining promising accuracies with the aid of a public heart failure dataset (best accuracy of 88.5%).

Abstract:
During the last years and especially during the pandemic more and more people have purchased different devices like smart watches for the purpose of monitoring their health from their home. While there are many devices that can help with remote health monitoring, most of them are only capable of monitoring one vital function or they are very expensive and difficult to use for some people. Considering this, the proposed project has the purpose of designing and implementing a device that includes multiple measurements for health monitoring in a single, more affordable, and easier to use system. The system can be used to measure pulse, SpO2 level, spirometry values like FEV1 and FVC or recording an electrocardiogram. The device also has a mobile application for support in which the results are displayed and interpreted.

Abstract:
Cognitive decline is a significant societal challenge, and the development of information and communications technology solutions to manage it is essential. The main challenges include the lack of comprehensive monitoring and assessment of cognitive decline, limited engagement with conventional cognitive stimulation applications, and the need for multi-domain interventions for cognitive function stimulation. In this paper, we present innovative technological solutions for managing more effectively the cognitive function of older adults experiencing mild cognitive impairment that integrates social robots, Internet of Things monitoring, and machine learning. It offers services for dealing with non-invasive monitoring and machine learning assessment of cognitive decline by fusing activity trackers and self-reporting data and personalized social robot-based interventions for cognitive function stimulation and social interaction facilitation using cognitive games and pleasant activities such as theatre therapy or storytelling. We have defined setup of the trials, procedures, and indicators to measure the impact on various care dimensions such as improving the quality of life and cognitive state of older people, Improving the quality of care and the cost efficiency or level of support offered for the informal or formal carers.