Robustness of Twitter Sentiment Classifiers Under Adversarial Text Perturbations in High-Volume Brand Interactions

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

Social media platforms, particularly Twitter, have become critical channels for brands to monitor consumer sentiment in real-time. While transformer-based models such as BERT and RoBERTa have achieved state-of-the-art performance in sentiment classification, their vulnerability to adversarial text perturbations poses significant risks for brand perception monitoring. This study investigates the robustness of Twitter sentiment classifiers under adversarial attacks in high-volume brand interaction contexts. Using a corpus of 2,282,912 English tweets containing brand-related keywords, we systematically evaluate the performance degradation of BERT-based sentiment classifiers under three adversarial perturbation strategies: gradient-based word replacement, synonym substitution, and character-level modifications. Our findings reveal that BERT-based classifiers experience accuracy degradation from 89.4% to 71.2% under moderate adversarial perturbations, with gradient-based attacks proving most effective at inducing misclassification. The CONV-GBERT hybrid architecture, integrating BERT-GPT fusion with convolutional feature extraction, demonstrates superior robustness, achieving 84.9% accuracy under adversarial conditions—a 4.9% improvement over standard BERT. We further identify that negative sentiment expressions, critical for brand crisis detection, are disproportionately vulnerable to adversarial manipulation, with recall dropping from 88.1% to 64.3%. These findings underscore the urgent need for adversarially robust sentiment analysis frameworks in brand monitoring applications and provide actionable insights for developing resilient NLP systems for social media intelligence.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.