Research On Adversarial Attacks For Video Object Tracking

With the continuous innovation of internet technology and video technology,the wide application of intelligent hardware,and the vigorous development of industries such as self-driving cars and drones,modern society has put forward higher requirements for video processing and analysis capabilities.As one of the important research directions of computer vision,video object tracking has a large number of applications in security,automatic driving,military and other fields,and these fields have extremely high requirements for the robustness of video object tracking.Therefore,it is of great significance to study the adversarial attacks of video object tracking.However,the current video object tracking adversarial attack algorithm is not yet mature.Most attacks mothed need to be based on the internal details of the video object tracking model to carry out effective attacks.However,in actual scenarios,the information of the model is often unknown,so it is difficult to implement;and the attack generalization ability is limited.At the same time,adversarial attack algorithms for video object tracking generally need to use a large number of video annotations to generate adversarial samples,which consumes a lot of manpower and material resources.In view of the existing problems,this project proposes a transferable black-box attack method for visual object tracking based on important features and an unsupervised adversarial attack method for video object tracking based on cycle consistency.The specific research content and implementation are as follows:(1)Aiming at the problem that adversarial attacks are difficult to implement in actual scenarios and the generalization ability is weak,a transferable black-box attack method for video object tracking based on important features is proposed.Design an important feature attack method based on gradient perception and a feature similarity attack method based on timing perception.Through the important features that are highly related to the tracking target and have universality,transferable adversarial samples are generated on the source model,so that other internal Models with unknown information conduct black-box attacks and improve the generalization of the attacks.At the same time,using the timing information of the video,according to the similarity of the video between adjacent frames,the attack is carried out by reducing the feature similarity between adjacent frames.In this way,the information of video space and time can be fully utilized to generate efficient and highly transferable adversarial examples.Experiments show that the performance of the tracking model attacked by this attack method is greatly reduced,which effectively improves the attack migration and attack effect.In OTB benchmark,the tracking success and precision of Siam RPN tracking model are reduced by 71.5% and 79.9% respectively.In the three benchmark of VOT,the accuracy of the Siam RPN tracking model decreased by an average of 5.1%,and the average overlap expectation decreased by an average of 48.5%.(2)Aiming at the problem that the adversarial attack of video object tracking requires large-scale video annotation,which is inefficient.The cycle consistency adversarial attack method and the context adversarial attack method are designed.We use the principle of cycle consistency,that is,the object tracking model’s forward tracking and backward tracking results of the video should be the same,an unsupervised cycle-consistent adversarial attacks for visual object tracking is proposed.Without the need to use video annotations,by breaking the cycle consistency principle of the object tracking model,the forward tracking and backward tracking of the object tracking model are made as inconsistent as possible to generate effective adversarial samples.At the same time,we also use the attack object surrounding context area information to propose a contextual attack method,attacking the object area and its surrounding context area at the same time,reducing its response score to attack.Experiments show that this method effectively attacks video object tracking models.In the three of VOT,this attack method reduces the accuracy of UDT tracking model by an average of 21.3%,and the average overlap expectation is reduced by an average of 62.6%.In the OTB100 benchmark,the success of the UDT tracking model drops by 34.2%,precision dropped by 48.7%.(3)Based on the research results of this subject and the application of the adversarial attack algorithm for video object tracking,and integrate the transferable black-box attack on video object tracking based on important features and the Unsupervised cycle-consistent adversarial attacks for video object tracking,an adversarial attack prototype system of video object tracking is designed and implemented.This thesis has 32 pictures,12 tables,and 83 references.