An adversarial text generation method at multiple granularity

Adversarial samples pose significant challenges to machine learning models, highlighting the need for robust defenses. This paper presents a novel approach for studying the influence of adversarial samples on machine learning models by proposing a multi-granularity text adversarial sample attack method under the constraint of semantic similarity. The proposed method combines word substitution at the word level and a short text generation model at the sentence level to generate target samples. By incorporating semantic similarity limits and part-of-speech restrictions during the adversarial text generation process, the resulting adversarial text exhibits higher semantic similarity. Aiming at the problem of the exponential increase in computational demand brought about by the introduction of the genetic evolution algorithm, we propose a dynamic parameter adjustment method to achieve effective pruning of subclass target groups. Experimental evaluations demonstrate that the multi-granularity attack method outperforms existing techniques such as TextBugger, PWWS, VIPER, and others. It achieves higher success rates in generating adversarial samples while maintaining a higher degree of semantic similarity. This research contributes to advancing the understanding of adversarial attacks in the context of text-based machine learning models and provides insights into enhancing the robustness of such models against adversarial samples.