Security Evaluation of Emojis in NLP Tasks

Emojis have become very popular in recent years, and they help analyze the meaning of texts. In many Natural Language Processing(NLP) tasks, such as sentiment analysis and question-answering systems, the ability to process emojis appropriately can significantly enhance the performance of these task models. However, neural networks are fragile and inherently vulnerable to various attacks, such as adversarial attacks and backdoor attacks, which can seriously reduce the usability of the models. There is no systematic study of the security issues that emojis may pose to different classification models. This paper presents the first systematic study on the security of emoji-based classification models. We analyze the security issues that may be raised by emoji processing methods, using adversarial and backdoor attack methods across different models and application scenarios. Our research shows emoji processing can increase model performance and raise security concerns. Our experiments show that most word-based attack methods equally apply to emojis. Given the widespread use of emojis and flaws in their processing methods, emojis could be a severe security threat to NLP tasks. In addition, we propose the characteristics of emoji-based attacks. Finally, we summarize the possible defense mechanisms based on existing research.