Privacidade em algoritmos de aprendizado de máquina: comparação de duas implementações de Regressão Linear usando Criptografia Homomórfica

Abstract

Data privacy is an essential issue to be addressed in the current era of complex data flows. Traditional anonymization methods are no longer sufficient to ensure individuals’ privacy. New solutions are emerging and are constantly evolving. Among the main approaches widely discussed in the literature and explored in this work, Federated Learning (FL), Secure Multiparty Computation (SMC), Differential Privacy (DP), and Homomorphic Encryption (HE) stand out. In this context, the development of Machine Learning (ML) algorithms capable of balancing the utility and privacy of data is imperative. While most works focus on the inference stage in ML, the ideal scenario should also address the training stage in the application of algorithms that preserve privacy. This work compares two implementations of the linear regression model in both stages, analyzing elements such as error and execution time, within the broader context of ensuring data privacy and utility balance.