Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Tracking Robustness with Auxiliary Adversarial Defense Networks

Zhewei Wu, Ruilong Yu, Qihe Liu, Shuying Cheng, Shilin Qiu, Shijie Zhou

TL;DR

This work tackles adversarial vulnerability in visual object tracking by introducing AADN, an auxiliary UNet-based defense preprocessor that sits before the tracker and is trained with Dua-Loss guided adversarial training. The method defends both the template and search branches in a plug-and-play fashion, without retraining the tracker, and achieves real-time performance (about 5 ms per frame). Extensive experiments on OTB100, VOT2018, and LaSOT show strong robustness to non-adaptive and adaptive attacks, plus good transferability to heterogeneous trackers such as SiamRPN, SiamMask, and Ocean. The approach offers a practical, scalable defense for real-time tracking in adversarial environments with minimal computational overhead.

Abstract

Adversarial attacks in visual object tracking have significantly degraded the performance of advanced trackers by introducing imperceptible perturbations into images. However, there is still a lack of research on designing adversarial defense methods for object tracking. To address these issues, we propose an effective auxiliary pre-processing defense network, AADN, which performs defensive transformations on the input images before feeding them into the tracker. Moreover, it can be seamlessly integrated with other visual trackers as a plug-and-play module without parameter adjustments. We train AADN using adversarial training, specifically employing Dua-Loss to generate adversarial samples that simultaneously attack the classification and regression branches of the tracker. Extensive experiments conducted on the OTB100, LaSOT, and VOT2018 benchmarks demonstrate that AADN maintains excellent defense robustness against adversarial attack methods in both adaptive and non-adaptive attack scenarios. Moreover, when transferring the defense network to heterogeneous trackers, it exhibits reliable transferability. Finally, AADN achieves a processing time of up to 5ms/frame, allowing seamless integration with existing high-speed trackers without introducing significant computational overhead.

Enhancing Tracking Robustness with Auxiliary Adversarial Defense Networks

TL;DR

This work tackles adversarial vulnerability in visual object tracking by introducing AADN, an auxiliary UNet-based defense preprocessor that sits before the tracker and is trained with Dua-Loss guided adversarial training. The method defends both the template and search branches in a plug-and-play fashion, without retraining the tracker, and achieves real-time performance (about 5 ms per frame). Extensive experiments on OTB100, VOT2018, and LaSOT show strong robustness to non-adaptive and adaptive attacks, plus good transferability to heterogeneous trackers such as SiamRPN, SiamMask, and Ocean. The approach offers a practical, scalable defense for real-time tracking in adversarial environments with minimal computational overhead.

Abstract

Adversarial attacks in visual object tracking have significantly degraded the performance of advanced trackers by introducing imperceptible perturbations into images. However, there is still a lack of research on designing adversarial defense methods for object tracking. To address these issues, we propose an effective auxiliary pre-processing defense network, AADN, which performs defensive transformations on the input images before feeding them into the tracker. Moreover, it can be seamlessly integrated with other visual trackers as a plug-and-play module without parameter adjustments. We train AADN using adversarial training, specifically employing Dua-Loss to generate adversarial samples that simultaneously attack the classification and regression branches of the tracker. Extensive experiments conducted on the OTB100, LaSOT, and VOT2018 benchmarks demonstrate that AADN maintains excellent defense robustness against adversarial attack methods in both adaptive and non-adaptive attack scenarios. Moreover, when transferring the defense network to heterogeneous trackers, it exhibits reliable transferability. Finally, AADN achieves a processing time of up to 5ms/frame, allowing seamless integration with existing high-speed trackers without introducing significant computational overhead.
Paper Structure (17 sections, 6 equations, 3 figures, 5 tables, 1 algorithm)

This paper contains 17 sections, 6 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Adversarial Attacks in Object Tracking
  4. Adversarial Training
  5. Proposed Method
  6. Problem Definition
  7. Dua-Loss
  8. Dua-Loss Guided Adversarial Training
  9. Experiments
  10. Experiment Settings
  11. Robustness on non-Adaptive Attacks
  12. Robustness on Adaptive Attacks
  13. Generalization on Clean Samples
  14. Transferability on Different Visual Trackers
  15. Speed of Different Defense Patterns
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Visualized results of defense performance on 3 sequences from OTB100. Better viewed in color with zoom-in.
  • Figure 2: The training pipeline for deploying the proposed AADN on the search branch.
  • Figure 3: Visualized search regions from OTB100 otb2015 dataset and their corresponding heatmaps. The columns represent the types of visualized images, including the original search region, the heatmap after being attacked by the CSA method csa, and the heatmap obtained based on defense samples.