FE - DI | Dissertações de Mestrado e Teses de Doutoramento
Permanent URI for this collection
Browse
Recent Submissions
- Anomaly Detection in Microservices Using Execution Traces and Graph Neural NetworksPublication . Martins, Sara Maria da Silva; Freire, Mário MarquesTTraditional anomaly detection methods are often based on manual analysis of logs and are insufficient for the complex, multi-layered interactions of microservices. The existence of multiple calls to and from the same services, following different execution paths, is difficult to accommodate in traditional, usually linear, methods. Anomaly detection plays a key role in maintaining systems health and complying with Service-Level Agreements (SLAs) by identifying unusual behavior that may indicate system failures or security vulnerabilities. This shortcoming of traditional methods can mean that faults are not detected in microservice systems. Failure to properly monitor a system can lead to serious and potentially damaging failures, thus, to find ways of detecting new types of anomalies is imperative. This research work proposes a Graph Neural Network (GNN) model, with different convolutional layers and activation functions to optimize performance, using the AIOps Challenge 2020 dataset for evaluation. The results show that simpler models with fewer hidden layers achieve the best performance, often exceeding the accuracy of 99.9%. The research concludes that anomaly detection methods should be tailored to specific systems and that GNNs offers a promising approach to dealing with the complexities of microservice architectures.
- Using Genetic Algorithms to Automatically Generate Unit TestsPublication . Magalhães, Manuel Ferreira; Pombo, Nuno Gonçalo Coelho CostaUnit testing is one of the most important activities during the lifecycle of a software product. In this operation, it is possible to perform tests for isolated individual units of software to test their overall functionality without external interactions and to verify if they correspond to the defined system functional requirements initially set in the software product’s lifecycle. The creation of unit tests can currently be done manually, with the use of human activity, or automatically, either through the use of tools, frameworks, artificial intelligence, or dedicated algorithms for their generation. Automation in this area of unit test generation shows great potential in terms of process efficiency and in reducing costs and development time. This area of study has a significant impact on software quality, since it directly affects the software lifecycle, where the development of major components depends strictly on the good development and testing of individual units. This study presents an automatic approach for the generation of unit tests from an individual code snippet using genetic algorithms. Additionally, benchmarks are also performed between the algorithms to assess the performance of the automated process and the quality of the generated unit tests.
- Implementation and Evaluation of QUIC protocol on Internet of Things Monitoring SolutionsPublication . Almeida, Luis Filipe Pereira de; Silva, Bruno Miguel Correia daIoT is a technology that ensures the possibility of significantly improving our lives by connecting hardware devices and everyday objects to the internet. Subsequently, these devices will be used to collect, transfer, and analyze vast amount of data, which can then be used to act on our environment or interact with us directly. The market for this technology is rapidly expanding in various sectors such as transportation, manufacturing, healthcare, among others. Due to this, IoT is also undergoing rapid evolution, requiring the development of fast, secure, and reliable communications. In an attempet to address the issues refered before, the use of QUIC in IoT technology is being explored as an alternative to currently used internet transport protocols. Although initially designed to operate in conjunction with HTTP/3, it has features that can benefit IoT, such as low latency compared to the TCP protocol, multiplexing that resolves head-of-line blocking found in TCP, and the use of a connection ID to enable connection migration. This dissertation explores the implementation of the QUIC protocol on IoT monitoring solutions. We aim to develop, implement and evaluate this implementation by comparison to other well known and used IoT-related protocols, such as, HTTP, UDP, MQTT and CoAP. This work explores the use of the QUIC protocol in a real-world scenario. This scenario will be the RuralTHINGS project, which focuses on creating an intelligent system for monitoring health-hazardous gases, such as radon gas and carbon dioxide, in residential areas. The main objectives are to discover the advantages and disadvantages of using the QUIC protocol, the scenarios in which it should be used, and how it compares to other similar protocols.
- Cross-Dataset Deep Fake DetectionPublication . Marques, José Pedro da Costa; Neves, João Carlos RaposoThis thesis presents an in-depth study of deepfake technology, focusing on generative techniques and the critical challenge of detecting such forgeries. The study delves into the technical foundations of deepfake creation, emphasizing the use of Generative Adversarial Networks (GANs), Diffusion Models, and Autoencoders. Alongside the generative processes, this research examines a variety of detection algorithms, stressing the importance of specific feature analysis and deep learning detectors in identifying sophisticated forgeries. A novel approach is proposed in this thesis, incorporating advanced facial feature analysis using spatial, frequency, and temporal landmark features, with Gated Multimodal Units for feature fusion, allied to a triplet attention mechanism, and transformers to leverage temporal dependencies. These methodologies are designed to improve the detection of deepfakes, addressing the growing sophistication of such technologies and their ability to bypass traditional detectors. The contributions of this thesis extend both to academic understanding and practical applications in enhancing the robustness of deepfake detection mechanisms, ultimately contributing to a safer digital environment.
- Massive Text Data VisualizationPublication . Voltoline, Giovana; Pais, Sebastião Augusto Rodrigues FigueiredoThis thesis focuses on the development of a specialized solution for the visualization of textual data. With the increasing volume of unstructured text in various fields of research, there is a growing need for effective tools that help users analyze and visualize such data. The primary objective of this research is twofold: first, to investigate the needs and challenges faced by researchers working with text data, and second, to design and implement a practical solution that meets those needs. The initial phase involves an extensive review of the literature, aiming to understand the current landscape of text data visualization, categorize researcher objectives, and identify the processes they follow. This research informs the selection of visualization types and the most suitable Python libraries to implement them. The final product is made accessible as both a Python package and an open-source repository on GitHub, providing a flexible tool that addresses the key challenges in text data visualization. This thesis not only contributes a functional tool for researchers but also highlights the importance of targeted solutions in handling and visualizing unstructured textual data.
- A Machine Learning-based Decision Support System for IoT Monitoring SolutionsPublication . Ladeira, Edgar Filipe Loureiro; Silva, Bruno Miguel Correia daGlobally, environmental incidents occur daily due to the emission of pollutant gases, many of which are the main causes of environmental disasters and premature deaths. The significant impact of these gases on public health is a major concern, especially in rural regions. Inhalation of these compounds often affects the daily lives and health of people living in these areas, often without their awareness of the risks involved. The goal of this project is to develop a decision support system powered by Artificial Intelligence (AI) and a robust platform for visualizing statistics on pollutant gases. Using an Internet of Things (IoT) ecosystem, the system aims to notify residents about environmental conditions and prevent any risks, including exposure to hazardous gases. Data is collected by digital sensors, aggregated by a microprocessor, and transmitted to a centralized database. This database feeds into the digital platform, which provides comprehensive and easy-tonavigate graphs, helping users to clearly understand pollution levels. These visualizations assist authorities and users in making informed decisions about environmental safety. This work aims to find the best Machine Learning (ML) model for this type of data (time series), in order to generate accurate predictions, focusing on the exploration of various ML techniques. The AI-based decision support system uses ML techniques to analyze the collected data and provide assessments of hazardous gas levels, as well as future forecasts of these levels, using unidirectional LSTM layers overlaid on bidirectional LSTM layers, an architecture that has shown the best results. In areas where pollutant gases such as radon are common, these forecasts enable the generation of timely alerts and recommendations to reduce health risks. Overall, this study contributes to the development of an innovative system through the analysis and processing of sensory data, such as pollutant gases and environmental variables. In a real-world context, the forecast is integrated into the RuraLTHINGS platform, providing system users with information to make more informed decisions about their indoor spaces, with the goal of promoting better living conditions.
- Bug Taxonomy Classification Using Machine Learning AlgorithmsPublication . Caldeira, Beatriz Isabel Trocas; Pombo, Nuno Gonçalo Coelho CostaBugs are a natural occurrence in the realm of software development. As society increasingly relies on software, the frequency of these occurrences has naturally increased as well, potentially leading to catastrophic consequences for products or businesses. To monitor and manage the bug resolution process, various bug tracking systems have been developed. These platforms enable teams responsible for bug analysis to view identified bugs, track the most common issues, and monitor their resolution status. However, the challenge lies in analyzing each of these reports. In large-scale products, hundreds of reports can be submitted daily through these platforms, whether by other developers, pen testers, or end users. End users, due to their likely lack of knowledge about software, system architecture, or other inherent processes, may submit incorrect or unclear reports. Addressing the need for developers to manually analyze and classify each of these reports (a complex and time-consuming task) through automation is an idea that has been explored by various researchers. This research aims to develop a comprehensive and detailed classification schema that can provide additional, automated information to report analysts, thereby facilitating part of their analysis process. To achieve this, the proposed solution is based on leveraging the high performance and capabilities of the BERT model, a model rooted on Transformer architecture, to classify these reports as accurately as possible, based on the textual descriptions within each report. A dataset was created following this schema, designed through a divideand-conquer approach, and multiple models were trained on each category. The results indicate that BERT can form the basis of a robust solution for this purpose, even when provided with a small amount of labeled data as input and despite the need for some refinement in the training process. The best results were obtained when both the title and the description were used together as input data for the model, with one model achieving an overall accuracy of 75%, and the lowest accuracy being 54.6%.
- Performance Analysis of End-To-End Autonomous Driving Systems in Varying Simulated ScenariosPublication . Morgado, Ângelo Miguel Rodrigues; Pombo, Nuno Gonçalo Coelho CostaIn recent years, the importance of autonomous driving has increased significantly due to technological advances and the changing needs of society. Autonomous driving promises safer, more efficient transportation, with the potential to reduce traffic congestion and accidents. However, achieving fully autonomous driving remains a challenge. One of the main limitations is that current autonomous vehicles cannot drive safely in all scenarios, especially in adverse weather conditions. Testing these scenarios in the real world is impractical and often unsafe, which makes simulation an essential tool for developing autonomous driving agents. Simulation environments allow for extensive testing of autonomous vehicles, providing valuable information and helping to improve the reliability and safety of autonomous driving systems. Most commercially available autonomous vehicles follow a modular architecture, where each phase, from information collection, to vehicle control is separated into different modules arranged individually from each other. However, there are studies that discuss a different architecture, end-to-end, where all these modules are combined into one and trained simultaneously. This project has two main objectives, the first is to develop a tool called CARLA-GymDrive, whose functionality is to allow the user to train a reinforcement learning agent in an environment compatible with the gymnasium library in the CARLA simulator; the second is to train agents with the two autonomous vehicle architectures end-to-end and modular, along with two learning algorithms, DQN and PPO. Trained agents are subject to a series of tests to comprehensively assess their performance and robustness. This study not only highlights the fundamental role of simulation in the development of reliable autonomous systems, but also acesses the potential of the end-to-end architecture.
- Implementação de Sistemas IoT na Indústria para Otimização e Controlo de Tanques de Água ContaminadaPublication . Neves, Pedro Gomes; Silva, Bruno Miguel Correia da; Simões, LuísO presente relatório descreve o desenvolvimento de um sistema IoT de monitorização de tanques de água contaminada ao longo do parque industrial do Grupo TESTA, com foco na eficiência e robustez. O objetivo principal do sistema passa por ter uma monitorização e um controlo contínuo dos tanques de líquido contaminado, utilizando sensores ultrassónicos ligados a microcontroladores para receber e processar os dados, para em seguida os enviar para uma plataforma IoT. Foi desenhada uma arquitetura geral do sistema de forma a assegurar, o máximo possível, o funcionamento fiável e o baixo custo de aquisição dos componentes, bem como a criação de dashboards para a visualização dos dados e alertas em casos de níveis críticos. Para além da monitorização, o sistema tenciona incorporar uma camada de inteligência via algoritmos preditivos, que permitem antecipar situações de risco e propor ações preventivas. Este trabalho destaca-se pela combinação de tecnologias emergentes na resolução de desafios industriais, contribuindo para uma gestão mais eficiente e sustentável dos recursos.
- Vision-Based Waste Detection for Industrial Sorting LinesPublication . Inácio, Sara Oliveira; Neves, João Carlos Raposo; Proença, Hugo Pedro Martins CarriçoThe increasing production of waste, driven by population growth, has created challenges in managing and recycling materials effectively. Manual waste sorting is a common practice, but it has proven insufficient to handle the large quantities produced and poses health risks to workers involved in the process. The presence of valuable materials in waste streams directed to landfills contributes to soil and water pollution, negatively impacting human health and ecosystems. With advances in deep learning and computer vision technologies, new solutions have emerged to automate and optimize waste sorting processes. In this context, this dissertation proposes a computer vision-based approach for the automatic detection of valuable waste materials in a municipal solid waste (MSW) sorting line. To support this work, a dataset was developed, collected from a Mechanical-Biological Treatment (MBT) facility, comprising eight waste categories. Several state-of-the-art object detection models, including Faster R-CNN, RetinaNet, TridentNet and YOLO, were trained and evaluated. The results demonstrated promising performance, achieving a mAP of 59.7% on the test set. This research and the developed dataset represent a contribution to the application of these technologies in the industrial sector, enhancing worker safety, increasing the efficiency of the recycling process, and supporting the development of sustainable solutions for waste valorization.