SSEAP: Semantic-Syntactic Enhanced Adversarial Repair for Universal Text Defense

Designing robust defense strategies against adversarial attacks is critical for enhancing the robustness of large-scale pre-trained natural language models. Existing defense methods can generally be categorized into proactive and passive defenses. However, these methods are often tailored to specific attack algorithms or introduce new perturbations during the adversarial removal process, resulting in insufficient accuracy and generalization. To address these issues, this paper proposes the Semantic-Syntactic Enhanced Adversarial Repair for Universal Text Defense (SSEAP), a framework designed to achieve both universal and accurate text defense. Specifically, SSEAP integrates semantic and syntactic constraints into a text-denoising model. The semantic consistency constraint is designed to handle character and word-level adversarial perturbations, such as synonym replacements and word order changes. The syntactic dependency constraint is primarily employed to repair structural adversarial attacks at the sentence level. By combining these two approaches, SSEAP effectively counters various types of adversarial attacks while maintaining higher defense accuracy. We conducted extensive experiments on three popular text classification datasets, successfully defending against seven different adversarial attack algorithms. The experimental results demonstrate that SSEAP significantly outperforms existing methods in terms of both universal and accuracy (with an average AUA improvement of 22.76%).