A* inspired context-aware adversarial attack with attention-based scores

The widespread use of deep learning models in Natural Language Processing (NLP) has raised concerns about their vulnerability to adversarial attacks. This paper examines word-level adversarial attacks, particularly greedy-based methods, which employWord Importance Ranking (WIR) to identify and replace key words with synonyms. However, these methods are limited because they operate in isolation, modifying individual words without accounting for the broader context. To overcome this limitation, we propose a novel hybrid algorithm inspired by the <i>A</i>* algorithm. Our approach combines a k-means-based ranking strategy to estimate immediate costs with an attention-based mechanism that captures the global impact of perturbations as future costs. Our approach is evaluated against state-of-the-art black-box and white-box adversarial methods, demonstrating its superior performance in generating high-quality adversarial samples across multiple metrics.