Adversarial machine learning

• Grant number: 23/14007-9
• Support Opportunities: Scholarships in Brazil - Program to Stimulate Scientific Vocations
• Start date: January 15, 2024
• End date: February 19, 2024
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Paulo Sergio Ramirez Diniz
• Grantee: Pedro Henrique Machado Zanineli
• Host Institution: Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa (COPPE). Universidade Federal do Rio de Janeiro (UFRJ). Ministério da Educação (Brasil)

Abstract

When we discuss the state of the art in machine learning, artificial neural networks (ANNs), a biological approach to algorithms can be understood as a reference in the area of deep learning. Seeking to learn complex patterns in data sets and perform tasks, ANNs have an input, intermediate, and output layer format, in which data will be received, processed, and displayed. Therefore, the effectiveness of the neural network will be highly dependent on the quality of the data to be processed, so that, if the information contains excess residue or contains damage, the process of finding existing patterns becomes difficult. Despite the problem, an important solution found was the conceptualization of adversarial learning, which corrupts purposefully data in order to strengthen the learning of the artificial neural network. (AU)