Usando técnicas de federated learning para aprimorar modelos de inteligência artificial no contexto das instituições públicas brasileiras

Abstract

Artificial intelligence models have become frequent in several areas of knowledge to resolve different problems efficiently. Due to this, many Brazilian Public Institutions have invested in AI solutions to improve and optimize their services. However, these institutions, mainly public safety organizations, use sensitive privacy data in their solutions. Thus, the use of this data is bureaucratic. Furthermore, each institution explores a simple scenario that makes the AI models biased. The data sharing between institutions could create general datasets with a better capacity to create more robust models. However, due to the nature of the data, this type of action is often unfeasible. Federated learning has gained space in the recent literature to enable the sharing of AI models safely. With this technique, it is possible to transfer knowledge from various models to create an improved version of them. Based on that, this work proposes using federated learning to create a safe environment for sharing AI models among Brazilian Public Institutions. In this work, five different federated learning algorithms have been experimented with three different machine learning algorithms. The experiments found that FedAvg, FedAdam, and FedYogi, together with the VGG16 and EfficientNet algorithms, achieve satisfactory performance in training classifiers in the context of Brazilian Public Institutions. Verifying that this result is maintained in scenarios with unbalanced or balanced datasets and many or few clients was possible.