Enhancing adversarial robustness in image classification through Winner-Take-All mechanism

Deep neural networks (DNNs) have achieved remarkable success in image classification across various applications. However, these models remain vulnerable to adversarial attacks, which can pose significant security risks, especially in safety-sensitive domains. Therefore, enhancing the adversarial robustness of deep learning models has become a critical research issue. Existing adversarial training methods based on Winner-Take-All (WTA) have made progress, but their approach of filtering large numbers of neurons leads to the loss of important information, thereby limiting the improvement of model robustness. To address this issue, we propose an improved adversarial training method, Winner-Take-All-based Adversarial Training (WTAAT). This method restricts the selection area of WTA to retain more important information, reduces neuron co-adaptation and mitigates the risk of overfitting. Additionally, we design a new trade-off loss function that balances robustness and accuracy during adversarial training, further enhancing the model's defense capability against complex attacks. Experimental evaluations on multiple benchmark datasets demonstrate that the WTAAT method achieves superior robustness against effective adversarial attacks.