https://repositorio.ufu.br/handle/123456789/18920 Sat, 04 Apr 2026 04:16:55 GMT 2026-04-04T04:16:55Z https://repositorio.ufu.br/handle/123456789/48556 Title: Aprendizado de redes neurais profundas para diagnóstico molecular rápido e não invasivo de COVID-19 pela saliva Abstract: The COVID-19 pandemic exposed limitations of conventional diagnostic methods, such as RT-PCR, which, despite their high sensitivity, are costly, time-consuming, and logistically complex. Vibrational spectroscopy emerges as a promising alternative, offering rapid, low-cost, and reagent-free analyses. However, the complexity of biochemical changes reflected in Raman and ATR-FTIR spectra demands advanced analytical methods that surpass traditional chemometric approaches in extracting nonlinear patterns. This work develops and evaluates models based on deep learning, specifically Convolutional Neural Network (CNN) architectures and Convolutional Neural Networks integrated with Long Short-Term Memory (CNN-BiLSTM), applied to the analysis of non-invasive biological sample spectra for COVID-19 detection. CNNs offer superior capability for extracting local spatial features from spectra through convolution operations, while BiLSTM networks complement this analysis by capturing bidirectional temporal dependencies in sequential spectral data, enabling a more comprehensive understanding of complex spectral patterns. Two architectures were developed: CNN-Spectra was designed to analyze Raman spectra from blood serum, achieving 96.8% accuracy; CNN-BiLSTM-Spectra was developed to analyze ATR-FTIR spectra from saliva, reaching an average accuracy of 80% and outperforming traditional and state-of-the-art models. The results confirm the effectiveness of the proposed models in extracting relevant features from complex spectral data, improving classification between positive and negative COVID-19 samples. It is concluded that the integration of deep learning with vibrational spectroscopy is a promising approach for biomedical diagnosis, contributing to faster, more accurate, and sustainable detection methods. Mon, 22 Sep 2025 00:00:00 GMT https://repositorio.ufu.br/handle/123456789/48556 2025-09-22T00:00:00Z https://repositorio.ufu.br/handle/123456789/48491 Title: Redes complexas para classificação de alto nível de assinaturas vibracionais moleculares Abstract: Attenuated Total Reflectance Fourier Transform Infrared (ATR-FTIR) spectroscopy is a promising technique for rapid and non-invasive diagnostics, yet its clinical application is often constrained by significant analytical challenges. The high dimensionality, complexity, and spectral overlap inherent in real-world data make it difficult to distinguish between different pathological conditions, a task where both conventional classifiers and deep learning models have shown limitations. To address these challenges, we propose ATLA (ATR-FTIR Topological Learning Analysis), a novel framework that integrates topological, structural, and physical information for enhanced spectral classification. Our methodology leverages the feature extraction capabilities of Complex Networks in synergy with traditional classifiers (Support Vector Machines and Linear Discriminant Analysis) and a state-of-the-art deep learning architecture (Convolutional Neural Network-BiLSTM). We assessed ATLA's performance on challenging real-world ATR-FTIR datasets for the diagnosis of COVID-19, Hepatitis B and D, and Chagas Disease. Our findings indicate that ATLA significantly improves predictive outcomes across key metrics, including accuracy, precision, sensitivity, specificity, and F1-score. Beyond predictive power, ATLA also enhances model interpretability by using SHAP (SHapley Additive exPlanations) to highlight the most relevant spectral bands and topological features driving the classification. Ultimately, ATLA emerges as a robust and effective approach, poised to elevate spectral diagnostics to a new level of accuracy and interpretability. Thu, 25 Sep 2025 00:00:00 GMT https://repositorio.ufu.br/handle/123456789/48491 2025-09-25T00:00:00Z https://repositorio.ufu.br/handle/123456789/48464 Title: Identificação de ameaças em fóruns da Dark Web e Surface Web: um estudo sobre a evolução temporal das discussões e generalização de modelos Abstract: In light of the increasing structuring of cybercrime and the rise of coordinated threats in anonymous environments, Cyber Threat Intelligence (CTI) becomes essential for proactive defense. This dissertation proposes two case studies aimed at evaluating practical applications of CTI through the integration of data mining techniques in Surface Web and Dark Web forums. The first study investigates the temporal evolution of discussions between 2015 and 2024 using topic modeling with LDA, identifying seasonal patterns and a thematic transition from technical debates to criminal practices, such as the commercialization of personal data in the Portuguese language. The second study evaluates the effectiveness of transfer learning to overcome the scarcity of labeled data in the security domain. To this end, a model based on the LightGBM algorithm with TF-IDF representation was employed, previously developed by a member of the same research project, and applied across distinct domains and multilingual environments, Portuguese and English. The results demonstrate that the model exhibits generalization capability by isolating risk-related vocabularies in new sources, such as Dark Web marketplaces and generic discussion forums, although it shows high sensitivity to technical terms. This study contributes to the development of CTI models with high efficiency and adaptability across heterogeneous data sources, supporting the anticipation of incidents even in the presence of limited labeled data. Sat, 28 Feb 2026 00:00:00 GMT https://repositorio.ufu.br/handle/123456789/48464 2026-02-28T00:00:00Z https://repositorio.ufu.br/handle/123456789/48334 Title: Recovering Chest X-RAY Images from Adversarial Attacks in Pneumonia Classification: Genetic Algorithm-based Adaptive Compression (GA-AC) Abstract: Adversarial attacks have become a critical threat to the reliability of deep learning models applied to medical image processing. Techniques such as the Fast Gradient Sign Method (FGSM), the Iterative Fast Gradient Sign Method (I-FGSM), also known as the Basic Iterative Method (BIM), and the Projected Gradient Descent (PGD) can induce significant prediction errors, compromising diagnostic accuracy and patient safety. To ad- dress this challenge, this study proposes an innovative approach called Genetic Algorithm- based Adaptive Compression (GA-AC), developed to recover images perturbed by various types of adversarial attacks. The GA-AC method employs a genetic algorithm to opti- mize image compression parameters, aiming to maximize the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM), thereby preserving essen- tial diagnostic features in the restored images. Experiments conducted on multiple X-ray images demonstrated the effectiveness of GA-AC, which successfully restored model per- formance after different adversarial attacks, recovering, for example, the F1-score from 24.14% to 98.10% after the FGSM attack, and achieving similarly robust results against I-FGSM, and PGD. Tue, 16 Dec 2025 00:00:00 GMT https://repositorio.ufu.br/handle/123456789/48334 2025-12-16T00:00:00Z