Boosting Robustness of Silhouette-Based Gait Recognition Against Adversarial Attacks

Gait recognition has found extensive applications in the field of biometric identification due to its advantages of long-distance and non-contact capabilities. Current research indicates that silhouette-based gait recognition exhibits rich features of human gait. It typically involves spatial feature extraction and gait temporal modeling, achieving widespread utilization with high accuracy. However, silhouette-based gait recognition is vulnerable to adversarial attacks. Adversarial examples are data samples generated by introducing minimal noise to original samples, imperceptible to human vision, yet deep neural networks are highly sensitive to them, leading to misclassification. In response, we propose a defense against adversarial attacks for silhouette-based gait recognition. This method constrains convolutional kernels to emulate the structure of edge detection operators, aiming to enhance edge information in images. The objective is to compel deep neural networks to focus more on semantic information in gait silhouette images and reduce feature deviations induced by adversarial perturbations. The method can significantly improve the adversarial robustness of silhouette-based gait recognition.