Research On The Robustness Of Siamese Network Tracking Algorithm Based On Adversarial Attack

Visual Object Tracking(VOT)is an important research branch in the field of computer vision.It has important application in tasks such as intelligent video monitoring and human-computer interaction.In recent years,with the wide application of deep learning in computer vision,more and more tracking algorithms based on deep learning have emerged.In this thesis,existing deep learning based VOT algorithms are surveyed,and we have found that the tracking algorithms based on Siamese network can achieve an acceptable tracking accuracy with reasonable processing speed.Nevertheless,our study has shown that deep neural network for VOT is also vulnerable to adversarial attacks.This observation has posed a huge challenge on the robustness of Siamese network trackers.In such a background,this work evaluates the degree of effectiveness of adversarial attacks on object tracking.It is the first study that applies the universal adversarial attack method to the fully-convolutional Siamese network(Siam FC)algorithm.We also study possible defense methods,which provides a guide for other researchers who work on the security of Siamese network based VOT algorithms.In this thesis we consider the following four aspects:(1)In the view of the potential security risks of adversarial attack in the field of computer vision,this paper presents the robustness of Siamese network tracker in different attack environments,considering the threat of adversarial attack based on deep learning target tracking algorithm.(2)In order to evaluate robustness of the Siam FC algorithm,this research applies one pixel attack and universal attack on the fully-convolutional Siamese network tracking algorithm.The experimental results show that the universal adversarial attack method has imposed a significant negative effect on the accuracy of the fully-convolutional Siamese network,which reduces the tracking accuracy of the Siam FC tracker by as much as 25.3%.The experimental results suggest that the fully-convolutional Siamese network tracking algorithm has poor robustness and is vulnerable to the universal attack method.(3)We apply adversarial training to obtain a new tracking model.The performance of the new model has achieved some improvement on the original dataset and the universal adversarialattack dataset.However such an improvement is not significant.(4)Combined with the characteristics of deep learning framework,we have proposed an improved algorithm called Siam FC-NK(Siam FC New Kernel)based on Siam FC algorithm.The new Siam FC-NK algorithm has significantly better performance on the original dataset,as well as better robustness on the attacked dataset.