|
|
Generování syntetických snímků duhovky
Haršaník, Dominik ; Juránková, Markéta (oponent) ; Vaško, Marek (vedoucí práce)
This thesis explores various methods for conditional generation of synthetic iris images using artificial intelligence. Current methods do not allow the generation of images based on different input parameters. I addressed this issue by employing an Auxiliary Classifier GAN (ACGAN) network, which utilizes generative adversarial model techniques while considering conditions such as the right and left eye, gaze direction, or scene lighting. In this thesis, I present a method capable of generating realistic synthetic iris images that also takes into account specified conditions and requirements. This work can be utilized in biometric identity verification, expansion of existing datasets, and training artificial intelligence algorithms for iris recognition.
|
|
|
Big Data Analysis Techniques for Network Traffic Monitoring: The Story of DNS over HTTPS Detection
Jeřábek, Kamil ; Laskov, Pavel (oponent) ; Wang, Peter Shaojui (oponent) ; Ryšavý, Ondřej (vedoucí práce)
Network monitoring plays a crucial role in the arsenal of tools used by network operators to ensure security. With the majority of network traffic now encrypted and the emergence of new protocols that extend encryption to previously unencrypted communications, traditional monitoring techniques that rely on the visibility of unencrypted network traffic have become obsolete. Consequently, solutions must now depend on the traffic metadata provided by widely used flow monitoring infrastructures. One of the protocols that get encrypted alternatives is DNS. DNS over HTTPS (DoH) is one of the attempts to encrypt DNS traffic that received broad adoption among users and resolvers. The~DoH implementation is already incorporated in most browsers, proxies, and operating systems. While DoH improves users' privacy, it leaves network operators and specialized Intrusion Detection Systems (IDS) blind to DNS traffic. Moreover, operators are unaware of DoH usage by users as DoH is designed to blend with other HTTPS traffic. Since its standardization in October 2018, the DoH has been studied extensively from various perspectives, including detection. This work proposes a reliable detection method using a combination of techniques, including machine learning, to identify DoH and distinguish it from regular HTTPS traffic, bringing awareness to network operators and allowing them to act according to their security policies. The work studies DoH thoroughly aligned with the data-centric concept of machine learning, enabling the creation of comprehensive datasets and designing effective practical detection mechanisms utilizing data sources of broadly present flow monitoring infrastructures. Moreover, the proposed detection method is tested in various scenarios, uncovering its characteristics and effectiveness compared with other state-of-the-art approaches.
|
|
|
Detection of Intensity in Sentiment Analysis of Czech
Dargaj, Jakub ; Tamchyna, Aleš (vedoucí práce) ; Mareček, David (oponent)
Postojová analýza sa zaoberá automatickou extrakciou subjektívnych informácií z textu. Cieľom práce je predpovedať intenzitu postoja v českých textoch. Na riešenie tejto úlohy sme pripravili dataset filmových hodnotení užívateľov Česko-Slovenskej filmovej databázy. Porovnávame niekoľko metód strojového učenia, pričom sa zameriavame na extrakciu číselných atribútov z textových dát. S využitím konvolučných neurónových sietí a korpusovo závislého trénovania vektorových reprezentácií slov sa nám podarilo prekonať základné modely a dosiahnuť presnosť podobnú najnovším výsledkom v tejto oblasti. V práci taktiež analyzujeme model logistickej regresie na porovnanie použitých jazykových prostriedkov medzi recenziami s rôznymi stupňami hodnotenia.
|
|
|
Comparative Analysis of Gaussian Process Regression Modeling of an Induction Machine: Continuous vs. Mixed-Input Approaches
Bílek, Vladimír
This paper investigates the application of machine learning technique for modeling continuous and mixed-input parameters of electrical machines. The design of electrical machines typically requires the consideration of certain parameters as integer values due to their physical significance, including the number of stator/rotor slots, stator wires, and rotor bars. Traditional machine learning methods, which predominantly treat input parameters as purely continuous, may compromise modeling accuracy for such applications. To address this challenge, models capable of handling mixed-input parameters were used for the case study. Two training datasets were generated: one with purely continuous inputs and another with both continuous inputs and a categorical parameter, specifically, the number of stator conductors. Gaussian process regression was employed to build three models: two with continuous kernels, trained on both datasets, and one with a mixed kernel, trained only on the dataset containing a categorical parameter. A comparative analysis, demonstrated on a 1.5 kW induction machine - though applicable to a wide range of machines - illustrates the differences between the proposed approaches. The results highlight the importance of selecting an appropriate model for the Multi- Objective Bayesian optimization of electrical machines.
|
|
|
Biometric fingerprint liveness detection
Rišian, Lukáš ; Vítek, Martin
This work addresses the problem of biometric recognition of fingerprint liveness to identify and differentiate between real fingerprints and their artificial replicas. The main objective was to identify the features that are crucial for fingerprint liveness recognition and based on these features to propose an efficient classification algorithm. We worked with the LivDet database from 2009, which contains both real and fake fingerprints. This database has been used in a worldwide competition and the results of all implemented algorithms are publicly available for subsequent comparison of success rates. An important part of this work was the preprocessing of the image data, which was crucial for testing the selected features and implementing the algorithms. We analyzed more than 180 different features from which we selected the most relevant ones. We then used the selected features to develop several fingerprint recognition and classification algorithms. Using the selected features, several possible variations of the algorithms have been proposed. Among all the implemented algorithms, we achieved the best result of almost 90%. Compared to other algorithms that have been implemented for the same purpose and have been used and tested on the same database, this can be considered a satisfactory and reliable result. In conclusion, the main objective of this work was to provide an efficient, secure, and reliable solution in the field of biometric fingerprint spoof detection.
|
|
|
Automating Antibiotic Susceptibility Testing with Machine Learning for Disk Diffusion Test Analysis
Lepík, Jakub ; Čičatka, Michal
Rapid and reliable antibiotic susceptibility testing (AST) methods are imperative in response to the escalating challenges of antimicrobial resistance. This study focuses on enhancing disk diffusion testing, a cornerstone of AST, by integrating machine learning and automation. Leveraging state-of-the-art object detection models, including EfficientDet and Mask R-CNN and image-processing approaches, our methodology addresses the need for standardized evaluation processes across diverse laboratory equipment while enabling the integration of mobile devices into the workflow, democratizing AST, and enhancing its accessibility. We utilize a comprehensive disk diffusion dataset for object detection models captured by devices like mobile phones and professional solutions. Additionally, our experiments lay the groundwork for a web application adopting a device-agnostic approach, promising improved accessibility and efficiency in AST analysis.
|
|
|
Mapping and analyzing of signal coverage of 4G/5G mobile networks
Baránek, Michal ; Polák, Ladislav ; Kufa, Jan
This paper addresses the enhanced measurement of signal coverage, capacity, and reliability in mobile networks, particularly with the growing prevalence of 4G and 5G technologies. Given the escalating importance of these networks in everyday activities, there arises a demand for open-source solutions to evaluate and enhance their performance effectively. The objective of this research is to analyze gathered data to pinpoint areas necessitating network enhancements and to develop opensource software and hardware solutions for extracting essential performance metrics (KPIs) from 4G/5G networks. The proposed system offers an interface for assessing network performance and signal coverage, enabling cost-efficient measurements across diverse environments.
|
|
|
Deep Learning for Agar Plate Analysis: Predicting Microbial Cluster Counts
Čičatka, Michal ; Burget, Radim
Manual analysis of agar plates remains a bottleneck in microbiology, hindering automation efforts. This study investigates the feasibility of using machine learning for automated microbial cluster count detection from agar plate images. We employed various methods, including elbow detection (baseline) and supervised learning models (Support Vector Regression, Simple CNN, XGBoost, Random Forest, pre-trained VGG, and pre-trained Inceptionv3). The results demonstrate that machine learning models significantly outperform the baseline, achieving lower prediction errors and higher accuracy in identifying the correct number of clusters. Notably, both pre-trained VGG and InceptionV3 achieved strong performance, highlighting the effectiveness of transfer learning for this task. InceptionV3 exhibited the lowest error rates overall. This study establishes a foundation for developing robust automated systems for quantifying microbial growth, potentially streamlining workflows and improving efficiency in microbiological research and clinical settings.
|
|
|
Machine Learning-based Fingerprinting Localization in 5G Cellular Networks
Dinh Le, Thao ; Mašek, Pavel
This study explores the viability of employing machine learning (ML)-based fingerprinting localization in 5G heterogeneous cellular networks. We conducted an extensive measurement campaign to collect data and utilized them to train three ML models: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM). The findings reveal that RF delivers the highest accuracy among the three ML algorithms. Furthermore, the results indicate that 5G New Radio (NR) can benefit the most from this localization method due to the dense deployment of base stations, achieving median localization errors of 17.5 m and 106 m during the validation and testing phases, respectively.
|
|
|
Detection of parking space availability based on video
Kužela, Miloslav ; Frýza, Tomáš
This paper deals with the use of Machine vision and ML (Machine Learning) for a parking lot occupation detection. It presents and compares an already existing technology that solves such a problem with an AI (Artificial Intelligence) usecase. It introduces tools used to train and create such models and their subsequent results as well as a dataset that was used to verify the trained networks and discusses the future of how such a technology could be used to effectively and more affordably detect occupied parking spaces on parking lots.
|
host ::
RSS.