Towards adversarial defense with multi-stage information fusion for visual tracking

In recent years, the effectiveness and stealth of adversarial attack methods have been continuously improving, posing significant challenges to the robustness and accuracy of visual tracking. Currently, most adversarial defense methods are focused on image classification tasks, while research on defenses in visual tracking is still in its early stages. To enhance the defense capabilities of current visual tracking methods against adversarial attacks, this paper proposes a multi-stage information fusion defense (MIFD). By combining defense strategies such as wavelet domain feature restoration, frequency domain low-pass filtering optimization, and non-local means denoising, MIFD effectively removes adversarial perturbations and enhances target-specific features. Additionally, MIFD introduces a residual fusion mechanism to adjust the fusion weights of denoising results, optimizing the balance between denoising and detail restoration through residual fusion. We validated the defense performance of MIFD against most black-box and white-box attack methods on six benchmark datasets. Experimental results demonstrate that MIFD significantly restores the tracking performance of the trackers, enhancing the robustness and accuracy of existing visual tracking methods under adversarial attacks.