Evolution of Word-Level Adversarial Attacks and its Categorization in Natural Language Processing

Natural Language Processing (NLP)-based applications widely use Deep Neural Network (DNN) models, significantly enhancing tasks such as Machine Translation, Text Classification and Question Answering. Though it successfully solves real-world tasks, it is vulnerable to adversarial attacks, where subtle modifications to input text can cause wrong predictions. Adversaries can create perturbations at the char, word and sentence levels. Among these, word-level attacks are a big concern because they are hard to detect. They alter words by inserting, deleting or replacing them without changing syntactic and semantic similarity. This can make models less accurate and reduce people’s trust in them. To build strong defenses, it is important to understand how these attacks works and how they have changed over time. Existing countermeasures struggle to counter word-level attacks effectively. Many methods lack generalizability and adversarial training remains computationally expensive. This paper provides a structured review of word-level adversarial attacks, tracing their evolution over the past decade and categorizing them based on attack strategy, specificity and attacker’s knowledge. Furthermore, key research gaps and challenges in defending are highlighted. Finally, the future directions for enhancing the robustness of NLP models against word-level attacks are addressed.