Research on Adversarial Sample Defense Mechanism against CNN

Convolutional neural networks (CNNs) have a wide range of applications in tasks such as image classification and target detection, but their susceptibility to adversarial sample attacks seriously threatens model reliability. In this paper, a lightweight adversarial sample defense method based on multilevel feature reconstruction and dynamic randomization is proposed. Through multi-scale feature fusion and channel attention mechanism, efficient suppression of adversarial noise is realized in the feature space. This study validates the effectiveness of the method on CIFAR-10 and ImageNet datasets. Experimental results demonstrate that the proposed method achieves significant robustness improvements of 11.3% and 13.3% against FGSM and PGD attacks, respectively, with negligible degradation in clean accuracy (≤0.5%). In addition, the inference efficiency is improved by 35.9% over traditional adversarial training. This method provides a lightweight solution for model deployment in security-critical scenarios such as medical image analysis.