TEXTUAL ADVERSARIAL EXAMPLE GENERATION USING BIGRAM UNIGRAM-SEMANTIC PRESERVATION OPTIMIZATION ALGORITHM

The vulnerability of Natural Language Processing (NLP) models to adversarial attacks remains a critical challenge in the field of cybersecurity and AI robustness. While deep learning models have achieved high performance in sentiment analysis, they are susceptible to subtle input perturbations that induce misclassification. This study presents the design and practical implementation of a web-based system (Proof of Concept) that automates the generation of textual adversarial examples using the Bigram Unigram-Semantic Preservation Optimization (BU-SPOF) algorithm. Rather than proposing a novel attack algorithm, our primary contribution is the architectural integration of a dual-source candidate generation strategy (WordNet and OpenHowNet) and a Probability Weighted Word Saliency (PWWS) mechanism to perturb input text while maintaining linguistic coherence. The system was evaluated against a Long Short-Term Memory (LSTM) sentiment classifier using the IMDB dataset.