GBDE: Granular-Ball Textual Adversarial Attacks With Differential Evolution

Abstract

Neural network models in natural language processing have demonstrated remarkable proficiency, but theyare still vulnerable to adversarial attacks, where minute and indistinguishable alterations can profoundly disturb their performance. Existing methods struggle to define synonyms accurately and rely on greedy traversal or gradient computation, making them difficult to deploy. Inspired by these, this paper proposes a novel text adversarial attack method, which constructs rich candidate adversarial examples by non-trivially combining the Granular-Ball computing and Differential Evolution, termed GBDE. Through iterative optimization guided by a meticulously crafted loss function, GBDE iteratively optimizes to generate effective and high-quality adversarial examples. Specifically, GBDE realizes a multi-granularity granular-ball clustering method following the data features, to establish robust synonym sets. It then performs synonym substitutions on target samples to create candidate adversarial samples and further diversifies these samples through a customized differential evolution process. Additionally, the loss function incorporates a reward and punishment mechanism to steer the optimization of attacks. Empirical results showcase GBDE’s superiority when contrasted with cutting-edge models. Notably, GBDE produces high-quality adversarial texts with a larger success rate. Besides, ablation studies confirm the contributions of granular-ball clustering and differential evolution in text attacks, and the rationality of the loss function design.