Uma arquitetura serverless para aplicação de aprendizado federado

Abstract

Nowadays, companies save their customer data and analyze them a posterior to benefit the company, to improve the service provided or understand the customer’s behavior to obtain, for example, a higher number of sales. This data analysis is usually performed using some machine learning technique. These techniques usually need a large amount of data to work satisfactorily, which can be a hindrance for some companies. Therefore, to take advantage of these techniques, it is necessary to consider collaborating with competing companies in the same industry and sharing data between them. However, there may be better strategies than this approach as it would result in sharing customer information with direct competitors. That said, federated learning can be used to build a predictive model in such cases. Federated learning enables collaborative training across devices without sharing customer data with a central server. This approach reduces the dependency on cloud resources since it transfers the responsibility to the user and improves data privacy compared to the traditional and centralized machine learning method. In this case, the models are trained on the users’ devices and then sent to a centralized server to perform model aggregation using a specific algorithm. This way, there is no sending of data from the users. Another benefit of using federated learning is that companies with a large amount of data can join together to generate an even more general model of their customers’ behavior. In recent years we have seen the emergence of a new architecture called serverless computing. In this new paradigm, developers can focus more on functionality and less on server configuration, making implementing or removing functionality easier and faster. Also, scalability problems are solved by the cloud provider automatically and quickly. This work proposes an architecture that uses serverless computing and federated learning techniques to build an application that is able to train models and predict new values with a high degree of accuracy. We performed experiments to evaluate our architecture in the cloud and edge computing environments. Our results show that it is possible to use serverless computing and federated learning and still obtain satisfactory prediction results in both environments.