Adversarial Robustness Evaluation of Traditional and Transformer-Based NLP Models for Toxic Comment Detection

Abstract

The increasing prevalence of toxic online content has created significant challenges for digital platforms, necessitating the development of automated and robust toxic comment detection systems. Traditional Natural Language Processing (NLP) approaches based on statistical text representations often achieve strong baseline classification performance but remain vulnerable to adversarial text obfuscation techniques designed to evade content moderation systems. This research investigates the effectiveness and adversarial robustness of two NLP classification approaches: a traditional TF-IDF Logistic Regression model and a transformer-based Bidirectional Encoder Representations from Transformers (BERT) model. Using the Jigsaw Toxic Comment Classification dataset, the problem was simplified into binary toxic versus non-toxic classification. To evaluate robustness, adversarial augmentations including leetspeak substitution, character deletion, and spacing attacks were applied to toxic comments. Experimental results demonstrated that the BERT model significantly outperformed the Logistic Regression baseline in both standard classification performance and resilience against adversarial perturbations. While the Logistic Regression model experienced substantial performance degradation due to its dependence on exact lexical matching, the BERT model maintained strong performance through contextual embeddings and subword tokenization mechanisms. Among the evaluated attacks, substitution based obfuscation produced the greatest reduction in transformer performance, indicating that adversarial NLP remains an ongoing research challenge. The findings highlight the limitations of traditional bag-of-words approaches and demonstrate the advantages of transformer-based architectures for cybersecurity-oriented content moderation systems operating in adversarial online environments.