Uma pesquisa sistemática sobre defeitos em sistemas de aprendizagem de máquina

Abstract

In recent years, the integration of systems supported by machine learning (SAAM) has expanded across various fields, driving technological advancements in critical sectors such as healthcare, finance, security, and transportation. However, as these systems become increasingly complex and widespread, significant challenges related to the reliability and robustness of implemented solutions also arise. Issues such as classification errors, performance failures, and vulnerabilities to adversarial attacks can compromise not only the integrity of these systems but also the safety and trust in automated decisions. In this context, understanding and mitigating these defects is crucial to ensuring the safe and effective advancement of machine learning technologies. This Thesis presents a systematic literature review on defects, errors, and failures that most compromise the integrity and performance of machine learning-supported systems (SAAM). The analysis focuses on identifying and categorizing the main reported issues, as well as the artifacts produced by the academic community in response to these challenges. The study covers a wide range of articles published in recent years, highlighting predominant trends in SAAM research, such as system robustness against adversarial attacks, algorithm optimization, and data set management. Based on a research and article selection methodology, issues are identified and classified into subgroups, revealing that topics like image classification, anomaly detection, and defenseagainst adversarial attacks are central to current concerns. Additionally, the study examines the artifacts generated by these studies, such as machine learning models, optimization algorithms, and data sets, which are essential to advancing research. The review not only maps the critical failures that can deteriorate SAAM but also highlights proposed solutions to mitigate these risks, providing a comprehensive view of research and development efforts in the field. Understanding these dynamics is essential for researchers, developers, and decision-makers seeking to explore or enhance the implementation of machine learning technologies in critical systems.