Avaliação de modelo de Machine Learning para classificação de informações sigilosas: um estudo dos dados do Governo Federal

Abstract

In recent years, Public Administration has made significant progress in modernizing government systems in alignment with the Digital Government policy. This movement encourages the adoption of disruptive technologies, such as automation and machine learning, which can enhance administrative processes' efficiency and improve public service quality. Data processing has become a central issue in this context, requiring compliance with the Access to Information Law (LAI) and the General Data Protection Law (LGPD). Furthermore, strict adherence to information security protocols is essential to safeguard sensitive data. This article aims to evaluate a machine learning model for classifying confidential information based on historical data from the Federal Government. To this end, the study is divided into three sections: the first presents the justification and motivation for the work, along with the theoretical framework, considering the advancements in the use of ICTs in public services; the second section addresses the experimental structure, detailing the methodology used for data collection and processing, as well as the experiments conducted; finally, the results, challenges faced, and future work are discussed. The Random Forest model demonstrated good performance in classifying confidential information, achieving an accuracy of 94.39%. Other metrics, such as the Precision-Recall Curve (PRC), were also analyzed. However, challenges related to class imbalance and the differentiation between confidentiality categories were identified, highlighting the need for future refinements in modeling, the collection of a more representative volume of samples, and feature selection.