Abstract: Adversarial example is one of the main attack methods against deep learning models, and models with adversarial defense capabilities suffer from degraded prediction performance on normal samples or even significant accuracy drops, making practical deployment challenging. Input preprocessing methods, while removing adversarial perturbations, lack semantic constraints and tend to alter key classification features (e.g., packet rates), impairing the classification performance of normal samples in network intrusion detection. Threshold-based methods, which rely on one-dimensional separation boundaries, cannot distinguish samples with similar feature values, substantially reducing defense effectiveness. This method employs Kolmogorov-Arnold Networks (KAN) to infer semantic representations of behavioral intent and integrates a diffusion process with a conditional autoencoder to selectively remove adversarial perturbations while preserving key discriminative features under semantic guidance. Experiments on multiple real-world datasets show that the method achieves an accuracy increase of over 13% without affecting the model’s original prediction performance, effectively defends against major adversarial attacks and offers substantial practical value.