ARG-Net: Adversarial Robust Generalized Network to Defend Against Word-Level Textual Adversarial Attacks

Natural Language Processing models have strong performance across various applications, although they are susceptible to manipulation by adversarial instances. A minor disturbance has the potential to alter the outcome of the deep learning algorithm. Humans find it difficult to detect this type of disturbance, particularly adversarial instances created using word-level adversarial attacks. Char-level adversarial assault can be countered using grammar detection and word recognition. The current word-level textual adversarial attacks rely on the substitution of synonyms, resulting in perturbed text that often maintains proper syntax and semantics. Defending against adversarial attacks at the word level poses more challenges. This study introduces a novel system called Adversarial Robust Generalized Network (ARG-Net) that aims to protect against word-level adversarial assaults. ARG-Net improves the model's performance by using both adversarial training and data perturbation techniques during the training process. The results of our tests on two datasets demonstrate that the model, which is built upon our framework, successfully mitigates word-level adversarial assaults. The defense success rate of the model trained using ARG-Net is greater than that of the previous defense approaches when tested on 1000 adversarial samples. Furthermore, our model exhibits superior accuracy on the standard testing set compared to current defense techniques. The accuracy is comparable to, or even surpasses, that of the conventional model.