Viral infections are a major global health burden, making early and rapid validation essential to mitigate their impact. Combining small-sample learning with AI technologies enables efficient acquisition of antiviral peptide (AVP) structures, advancing antiviral research. We propose a generative model integrating 3D sample information and a dual-feedback mechanism for small-sample AVP classification. First, a structural enhancement module reconstructs 2D representations of AVP structures using PCA and denoises them via a simplified IFAN network. Next, the generative model, enhanced by the Observer-Discriminator Dual Feedback Mechanism (ODDFM), optimizes parameters to improve the quality and authenticity of generated AVP sequences. Finally, the generated structures augment small-sample data for AVP acquisition. On the AVP dataset, the model achieved an AUC of 0.963, precision of 0.955, recall of 0.966, and F1 score of 0.961, outperforming direct classification on the original dataset by over 0.05 in key metrics.