Research On The Anti Attack And Defense Scheme Of Gait Recognition

Gait recognition has the advantages of long-distance recognition and does not require the active participation of the recognized person,so it has a good application prospect.At present,many gait recognition models based on deep neural networks have achieved relatively high accuracy.However,many studies have shown that deep neural networks are vulnerable to adversarial attacks.By adding small perturbations to input samples,deep neural networks can recognize error.Therefore,it is a very important task to study the security of gait recognition model in the face of adversarial attacks.In this paper,the GAN-based adversarial sample attack method is used to attack the target gait recognition model,and the adversarial defense method is used to improve the security of the gait recognition network.The following results were achieved:(1)A GAN-based semi-white-box adversarial attack method is proposed.The method generates adversarial perturbations by using the GAN generator,and designs the loss function in the training process,so that the GAN generator can learn the distribution of the adversarial perturbations from the input samples.The adversarial perturbations are then added to the original samples to generate adversarial samples.Through experiments,it is found that the adversarial samples generated by this method are of higher quality and are difficult to be detected by the naked eye.(2)A perturbation constraint algorithm in the task of adversarial sample generation is proposed.The algorithm can constrain the GAN’s generator through a loss function to generate adversarial perturbations of a specific shape.In a physical environment,such attacks against the target model for adversarial perturbations of a specific shape are more feasible.(3)Improve the security of the gait recognition model in the face of adversarial attacks.The adversarial samples and the original samples are mixed according to a specific ratio to form a new data set,and then the gait recognition model is trained using the new data set,so that the target model learns the feature distribution of the adversarial samples.This method can effectively improve the performance of the gait recognition model.robustness.By training a binary classification model with original samples and adversarial samples,some adversarial samples can be detected,reducing the possibility of the gait recognition model being attacked by adversarial samples.