Abstract: While NLP models have achieved state-of-the-art performance across various classification tasks,their vulnerability to adversarial attacks remains a significant challenge.This paper introduces a novel retrieval-augmented classification approach designed to enhance model robustness against such attacks.By leveraging KNN retrieval mechanism,this method interpolates the predicted label distributions with those of retrieved instances,strengthening the model’s decision-making process in adversarial settings.A key innovation of this work is the decoupling of the representation spaces used for classification and retrieval,which mitigates performance degradation and training instability caused by shared representations.The proposed method is evaluated across a range of benchmark datasets under various adversarial attack scenarios,demonstrating substantial improvements in model robustness.Specifically,the accuracy drops typically observed under adversarial conditions are reduced by 30 percentage points to 40 percentage points,with the proposed approach maintaining performance stability even under intense attacks.Comprehensive experiments validate the effectiveness of the proposed method,highlighting the impact of both retrieval-augmented classification and decoupled representations in creating more resilient and reliable systems.

Key words: Adversarial defense, Retrieval-augmented classification, Natural language processing, Model robustness, KNN retrie-val, Representation learning