D-NEXUS: Defending text networks using summarization

Abstract

Sentiment analysis is an important tool for understanding consumer sentiment in e-commerce platforms. Usually, it is performed using Deep Neural Networks (DNNs), owing to their strong predictive and generalization capabilities. Unfortunately, DNNs are prone to adversarial attacks, which involve introducing imperceptible changes in the data with the deliberate motive of “fooling” the target model. It can lead to far-reaching consequences and pose an alarming issue to the credibility of e-commerce platforms. The existing text-based defenses, such as spelling correction and adversarial training, are largely ineffective against state-of-the-art adversarial attacks, most of which deal in word replacement, insertion, and deletion. We introduce an effective transformation-based defense strategy, D-NEXUS (DefeNding tEXt networks Using Summarization). It overcomes the drawbacks of existing defenses by summarising the input text before feeding it into the target model. Our extensive experiments on publicly available datasets show that D-NEXUS successfully defends against state-of-the-art attacks, in a time-efficient manner. © 2022 Elsevier B.V.