Research On Adversarial Attack Methods Under Twin Network Target Trackin

As an important research branch in the field of computer vision,object tracking plays a crucial role in ensuring public safety.With the rapid development of deep learning technology,object tracking algorithms based on siamese networks have shown great potential.They not only achieve remarkable accuracy but also solve the problem of slow online updating,becoming the mainstream method in the field of object tracking.However,existing research shows that object tracking models are vulnerable to adversarial samples,which poses a significant challenge for the deployment of deep object tracking models in the real world.Therefore,this thesis studies the adversarial attack methods of object tracking,improves the existing methods’ limitations,and makes the tracker difficult to accurately locate the target’s position.The research content of this thesis is as follows:(1)Addressing the limitation of existing visual object tracking adversarial attack methods that require online iteration,which is difficult to meet the real-time tracking requirements and cannot cause rapid drift in the center of the activated target,this paper proposes a multi-level loss object tracking adversarial attack method based on spatial awareness.First,starting from the semantic level,we design deception and drift losses to interfere with the tracker’s discriminative and regression abilities,making the predicted box tend to be in the background area and rapidly drift.Second,starting from the texture of the bottom-level feature,we design feature loss,which integrates spatial context and channel attention mechanisms to destroy the internal structure of the target in the feature space.Then,we use the proposed losses for joint training to obtain a generator based on the target search region.Finally,we use the generator to generate perturbations,transform clean samples into adversarial samples that can disrupt object tracking tasks,and cause the target’s motion trajectory to deviate from the original target center,thereby reducing tracking performance.The proposed method achieved a success rate decrease of 54% and an accuracy decrease of 70%on the OTB dataset,thereby achieving rapid and effective attacks on targets in complex scenarios.(2)Addressing the limitation of existing visual object tracking adversarial attack methods that require producing corresponding individual adversarial perturbations for each image,which consumes additional computational resources and has weak transferability,this paper proposes a universal robust object tracking adversarial attack method under unknown video conditions.First,we use Generalized Intersection Over Union(GIoU)as the metric to compare any two boxes and use the GIoU score between the original prediction output and the tracker’s predicted box as a guide.We perform flip attacks on the classification stage to confuse the target with the background area and perform scale attacks on the regression stage to reduce the overlap between the predicted box and the original target.Second,we combine the two attack methods to generate a universal adversarial perturbation.Then,we apply various transformation operations to the perturbation to improve its robustness.Finally,we add the universal adversarial perturbation to any unknown video image to generate an adversarial sample.The proposed method is capable of inducing the tracker to predict the motion of the target along the corner position direction,and has achieved good attack effects on several datasets.