ISEL - Eng. Elect. Tel. Comp. - Dissertações de Mestrado
Permanent URI for this collection
Browse
Browsing ISEL - Eng. Elect. Tel. Comp. - Dissertações de Mestrado by Field of Science and Technology (FOS) "Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática"
Now showing 1 - 3 of 3
Results Per Page
Sort Options
- Aplicações de modelos de linguagem de grande escala na cibersegurançaPublication . Conceição, Tiago Miguel Pestana; Cruz, Nuno Miguel MachadoA crescente complexidade e sofisticação das ameaças no ciberespaço têm impulsionado a procura por soluções inovadoras e eficientes no campo da cibersegurança. Neste contexto, conduziu-se uma investigação com o objetivo de avaliar a viabilidade de Large Language Models (LLMs) no que concerne à automatização da geração de código e configurações no âmbito da cibersegurança. A investigação centrou-se em mecanismos de ciberdefesa e aplicações de educação em cibersegurança, com particular ênfase em soluções de geração de honeypots, malware e exercícios de Capture The Flag (CTF). Foram avaliados sete modelos, incluindo o GPT-4, Gemini Pro e Claude Opus 3. A metodologia de avaliação assentou no desenvolvimento de dois mecanismos de avaliação, sendo o primeiro um novo benchmark Cybersecurity Language Understading (CSLU), baseado no Massive Multitask Language Understanding, constituído por questões de escolha múltipla sobre diversos domínios do conhecimento. As prompts foram concebidas com o intuito de avaliar o estado de conhecimento de cada modelo relativamente aos tópicos supracitados. O segundo mecanismo consistiu na avaliação da consistência, criatividade e adaptabilidade dos modelos referente à geração de artefactos. Os resultados evidenciaram uma notória proeminência referente ao tópico de malware, com quatro destes a alcançarem a pontuação máxima. Por outro lado, o desempenho na tarefa de CTF revelou uma maior variação de resultados. De um modo geral, os modelos GPT- 4, Gemini Pro e Claude 3 Opus demonstraram resultados consistentemente superiores entre os modelos estudados. Num segundo momento, pretendeu-se desenvolver uma ferramenta baseada na web, com o objetivo de fornecer uma prova de conceito dos estudos anterior realizados. A referida ferramenta, recorrendo aos melhores LLMs estudados, permite ao utilizador criar e lançar automaticamente serviços de segurança, como os mencionados honeypots ou exercícios de CTF. De uma perspetiva global, estas descobertas sugerem que a aplicação de LLMs em atividades de cibersegurança pode ser altamente vantajosa.
- Bitcoin Anomaly Detection (BAD) - use of machine learning for fraudulent transaction detectionPublication . Gomes, Nuno Gonçalo Rodrigues Cabral; Ferreira, Artur JorgeAbstract The increasing adoption of cryptocurrencies, especially Bitcoin, has significantly altered the financial landscape, enabling decentralised and pseudonymous transactions. While these characteristics foster innovation, they also present serious challenges in detecting fraudulent behaviour, including money laundering and investment scams. This dissertation introduces Bitcoin Anomaly Detection (BAD), a hybrid machine learning framework designed to detect anomalous transactions on the Bitcoin blockchain. The methodology integrates supervised and unsupervised learning techniques, applied to the Elliptic dataset comprising over 200,000 transactions with temporal and graph-based features. Feature engineering, class imbalance handling (e.g., SMOTE + ENN), and dimensionality reduction (PCA, UMAP) are employed. Several models are evaluated, including XGBoost, RF, and GNN, achieving up to 97.5% accuracy, 96.2% recall, and a false positive rate below 4.5%. Graph analysis revealed that illicit transactions tend to form sparsely connected “sink” nodes—receiving many inputs (high in-degree) but sending no outputs (zero out-degree)—a pattern typical of laundering. Semi-supervised learning and association rule mining were used to label unknown data and enhance classification reliability. The final pipeline combines feature selection, class instance sampling, and a hybrid semisupervised learning approach—bridging supervised and unsupervised methods—to classify transactions with high accuracy and robustness. Key challenges addressed include the scarcity of labelled data, severe class imbalance, and the evolving nature of fraud techniques. XAI components were incorporated to ensure interpretability and compliance with regulatory frameworks such as Markets in Crypto-Assets (MiCA) and Financial Action Task Force (FATF). The findings from our experimental evaluation demonstrate the viability of adaptive, interpretable AI solutions for safeguarding decentralised financial ecosystems and supporting efforts in Anti-Money Laundering (AML) and law enforcement.
- Drug recommendation system based on symptoms and user sentiment analysis (DRecSys-SUSA)Publication . Pinto, Ana Sofia Simões; Pato, Matilde Pós-de-Mina; Datia, Nuno Miguel SoaresAbstract The rapid growth of user-generated content on multiple online platforms has opened opportunities for improving decision-making across various domains, including healthcare. This dissertation focuses on the development of our Drug Recommendation System based on usergenerated content (DRecSys-SUSA), designed to assist healthcare professionals and patients by providing personalized drug recommendations and supporting informed decision-making. Our research leverages the UCI ML Drug Review dataset as the foundation for developing an advanced recommendation system. Our solution utilizes a combination of modern AI techniques, including Exploratory Data Analysis (EDA), data pre-processing, sentiment analysis (SA), and text generation using a fine-tuned Large Language Model (LLM). We design and propose a recommendation system framework, within which we implement multiple variants of DRecSys-SUSA using different combinations of AI techniques. Each variant generates medically relevant suggestions to user-specific inputs such as age, symptoms, and current medications. Through an iterative process of implementation and evaluation using an LLM-as-judge methodology with AI-generated real-world scenarios, we identify which AI techniques are most beneficial for providing clinically appropriate and user-friendly drug recommendations. The resulting insights contribute to the advancement of AI-driven healthcare tools by establishing effective approaches for leveraging user-generated content in medical recommendation systems.