Bernardo Lacerda Salgueiro Faria
Departamento de Engenharia Eletrônica e Computação (DEL), Escola Politécnica (EP), Universidade Federal do Rio de Janeiro, RJ
João Baptista de Oliveira e Souza Filho
Programa de Engenharia Elétrica (PEE) e Departamento de Engenharia Eletrônica e Computação (DEL), COPPE/EP, Universidade Federal do Rio de Janeiro, RJ
Adversarial Attacks, Machine Learning, Deep Learning, Passive Sonar Systems, Defense
Abstract
Passive sonar systems play a crucial role in submarine operations, serving as a fundamental tool for identifying potential threats by capturing, monitoring, and analyzing underwater noise. A wide range of machine learning models, particularly neural networks, have been successfully explored to automate this task. However, dissuasion tactics may compromise the performance of such systems - an issue not commonly addressed in the literature. This article proposes a simplified version of the Defensive Distillation approach, aimed at enhancing the robustness of neural classification models when subjected to dissuasion strategies based on adversarial attacks. Results obtained using noise emitted by 28 real ships from 8 classes, recorded in an acoustic range by the Brazilian Navy, confirm that the proposed technique allows for the cost-effective mitigation of adversarial attack effects while significantly reducing the associated computational effort.
