Monserrat Vázquez-Hernández

Analyzing students’ opinions about classes, instructors, and the overall learning environment enables institutions to understand students’ needs and propose improvements to enhance the learning process. Furthermore, analyzing opinions at aspect-level allows a detailed analysis and helps to identify the relevant aspects that require attention to enrich the students’ experience. Currently, opinion analysis systems implement deep learning models to increase precision. However, these models can be misled by strategically modified inputs, known as adversarial examples, producing incorrect classifications. In this work, we analyze adversarial examples and their behavior in aspect-level sentiment analysis models to design aspect-aware modifications. Specifically, we define a model for generating aspect-level adversarial examples that modify inputs by targeting aspect–term relationships. The proposed model reduces accuracy by 20.90%, highlighting vulnerabilities in aspect-level models, where even small and targeted changes can degrade performance, providing a foundation for the development of defense mechanisms to ensure reliable predictions.

Adversarial attacks deliberately modify deep learning inputs, mislead models, and cause incorrect results. Previous adversarial attacks on sentiment analysis models have demonstrated success in misleading these models. However, most existing attacks in sentiment analysis have applied a generalized approach to input modifications, without considering the characteristics and objectives of the different analysis levels. Specifically, for aspect-based sentiment analysis, there is a lack of attack methods that modify inputs in accordance with the evaluated aspects. Consequently, unnecessary modifications are made, compromising the input semantics, making the changes more detectable, and avoiding the identification of new vulnerabilities. In previous work, we proposed a model to generate adversarial examples in particular for aspect-based sentiment analysis. In this paper, we assess the effectiveness of our adversarial example model in negatively impacting aspect-based model results while maintaining high levels of semantic inputs. To conduct this evaluation, we propose diverse adversarial attacks across different dataset domains, target architectures, and consider distinct levels of victim model knowledge, thus obtaining a comprehensive evaluation. The obtained results demonstrate that our approach outperforms existing attack methods in terms of accuracy reduction and semantic similarity, achieving a 65.30% reduction in model accuracy with a low perturbation ratio of 7.79%. These findings highlight the importance of considering task-specific characteristics when designing adversarial examples, as even simple modifications to elements that support task classification can successfully mislead models.

In recent years, the use of deep learning models for deploying sentiment analysis systems has become a widespread topic due to their processing capacity and superior results on large volumes of information. However, after several years’ research, previous works have demonstrated that deep learning models are vulnerable to strategically modified inputs called adversarial examples. Adversarial examples are generated by performing perturbations on data input that are imperceptible to humans but that can fool deep learning models’ understanding of the inputs and lead to false predictions being generated. In this work, we collect, select, summarize, discuss, and comprehensively analyze research works to generate textual adversarial examples. There are already a number of reviews in the existing literature concerning attacks on deep learning models for text applications; in contrast to previous works, however, we review works mainly oriented to sentiment analysis tasks. Further, we cover the related information concerning generation of adversarial examples to make this work self-contained. Finally, we draw on the reviewed literature to discuss adversarial example design in the context of sentiment analysis tasks.