Javam C. Machado

A pesquisa proposta tem como objetivo avaliar Modelos de Linguagem de Grande Porte (LLMs) na tarefa de análise de resultados obtidos a partir de testes de segurança em ambientes online. O foco central está relacionado à interpretação automatizada de relatórios de exploração gerados por ferramentas OSINT, utilizados para identificar vulnerabilidades e, assim, reduzir a dependência de análises manuais exaustivas. O estudo aborda a lacuna associada à complexidade da inspeção desses relatórios, cuja natureza altamente técnica dificulta a identificação, a classificação e a priorização de falhas de segurança. A proposta também é motivada por diretrizes recentes da OWASP Foundation, que destacam a necessidade do uso de LLMs confiáveis e seguros em aplicações críticas. Para a realização da classificação de vulnerabilidades nos ambientes online analisados, é proposta uma taxonomia de integração unificada composta por cinco categorias Macro, construída a partir da integração de metodologias consolidadas dos frameworks OWASP Top 10 e FIRST CVSS. Os resultados demonstraram que modelos como qwen3.6-35b-a3b e gemma-4-26b-a4b apresentaram os melhores desempenhos gerais na avaliação, destacando-se nas métricas de F1-Score, Recall e Precision. Além disso, modelos compactos como o gemma-4-e2b também obtiveram resultados competitivos, indicando que LLMs menores podem representar alternativas viáveis em cenários com restrições computacionais.

Answering an unbounded number of multi-dimensional range queries while preserving privacy is a significant problem that has been the focus of recent studies since range queries serve as a core component of many data analysis tasks. Existing techniques that use local differential privacy (LDP), which adds noise to users’ data locally, make assumptions that significantly affect utility. In particular, they assume that all attributes in a dataset are correlated and materialize all pair-wise grids/trees. Moreover, they assume that each pair-wise correlation is equally relevant when answering queries. These assumptions result in a low signal/noise ratio, especially for datasets with a large number of attributes. To address the deficiency of existing techniques, we propose a new approach to answer multi-dimensional range queries that leverages multi-dimensional correlations and workload characteristics to selectively choose the data that is collected under the LDP model. Our approach first leverages the availability of a small number of DP users to build a correlation model to determine how attributes are correlated. Then, we use the notion of multi-dimensional grids to capture multi-dimensional correlations. That allows us to materialize fewer grids, boosting the signal/noise ratio. Finally, we propose a new optimization step that makes use of available workload characteristics to identify the user distribution that minimizes the total error when answering multi-dimensional queries among users during the grid assignment process. We conduct extensive experiments on real and synthetic datasets and demonstrate the superiority of our method over existing approaches.

Background: Traditional breast cancer screening protocols apply uniform strategies across populations with heterogeneous risk profiles. Leveraging routinely collected electronic health records to improve individual risk stratification has the potential to optimize early detection and better allocate clinical resources.Methods: We developed and validated a population-scale model to estimate individual breast cancer risk after negative or probably benign mammography examinations using longitudinal clinical data from the largest private healthcare system in Brazil, collected between January 2016 and December 2025. We selected BI-RADS 1, 2, and 3 yielding approximately 4.6 million examinations from 1.8 million females (median age 44 years [IQR 35--53]) from more than 200 hospitals and clinics in all five main regions of Brazil. Unlike image-based approaches, the model relied exclusively on radiologist report text and structured clinical data.Findings: The model predicted the occurrence of breast cancer within 5 years with an AUROC of 0.86 (95% CI: 0.85--0.87) and a concordance index of 0.86 (95% CI: 0.85--0.87), with similar performance across shorter prediction horizons. Flagging only the top 3.4% risk percentile identifies 50.1% (95% CI: 48.4%--52.2%) of exams from patients who would develop breast cancer within 5 years. In contrast, in the bottom 47% risk percentile, only 10.0% (95% CI: 8.98%--10.9%) develop cancer in 5 years. An explainability analysis identified surgical history, calcifications, and architectural distortions as the strongest drivers of high-risk predictions, while reassuring radiological findings and other clinical factors were prominent in low-risk predictions.Interpretation: This study demonstrates the feasibility of a state-of-the-art breast cancer risk model based on longitudinal text reports and structured clinical data, rather than raw images. Extracted from a large-scale, diverse, real-world health system, these results offer a practical and scalable tool for enhanced individual risk stratification.

Mobile devices have revolutionized image acquisition, enabling diverse communication and image-processing applications. One promising application is automated damage detection in printed circuit boards (PCBs), crucial for quality control in electronics manufacturing. However, unlike controlled environments, mobile device image acquisition introduces challenges such as non-uniform lighting, background interference, and varying camera resolution, which can affect the accuracy of deep learning models. This paper presents a case study investigating the impact of domain-specific, no-reference image quality metrics on the performance of deep-learning object detection models for PCB damage detection. We evaluate nine metrics, including five novel contributions, using a semantic segmentation approach to measure foreground and background quality. Our study assesses how these metrics influence the performance of deep neural network architectures, determining optimal thresholds that separate high and low-quality images. Experiments are conducted on real-life images captured with smartphones in industrial settings. Our findings indicate that filtering poor-quality images based on these metrics could significantly improve detection performance, offering practical benefits for mobile damage detection applications.

Mecanismos de seleção diferencialmente privados (DP) obtem o elemento de maior qualidade em um conjunto finito, garantindo forte privacidade. No entanto, a maioria dos métodos existentes usa sensibilidade global, frequentemente adicionando ruído excessivo e prejudicando a utilidade dos resultados. Propomos o algoritmo Smooth Noisy Max (SNM), que emprega sensibilidade suave para obter limites de erro mais precisos e maior acurácia sob condições moderadas. Experimentos mostram que o SNM supera os melhores métodos de seleção com DP em três aplicações: seleção de percentis, árvores de decisão gananciosas e florestas aleatórias.

Large and medium-sized Language Models (LLMs and MLMs) are fundamental in Natural Language Processing (NLP) applications but still present challenges in fairness evaluation, bias detection and mitigation. This work compares bias evaluation metrics such as SEAT, Adapted WEAT, and CEAT on models like BERT and LLaMA-2 and explores the automatic attribute selection by LLMs. The results show that automatically generated attributes enhance bias detection and that alternative metrics to cosine similarity significantly influence the outcomes. CEAT stands out in capturing intersectional biases, contributing to the development of more robust methods for fairness evaluation.

This work proposes a new operator, Purpose Scan (PS), which is introduced into the execution plan to ensure that data owners' consents for specific purposes are respected. It also introduces the Differential Purpose Scan, which incorporates differential privacy to prevent information leakage. Experiments show that PS improves performance compared to view-based approaches.