Blinding and blurring the multi-object tracker with adversarial perturbations.

Area of Science:

• Computer Vision
• Deep Learning
• Adversarial Machine Learning

Background:

• Deep multi-object trackers integrate detection and association, making them vulnerable to adversarial attacks.
• Existing attacks primarily increase ID switching, insufficiently degrading tracker performance.
• Imperceptible perturbations can deceive deep learning models, including object trackers.

Purpose of the Study:

• To propose a novel adversarial attack method, Blind-Blur Attack (BBA), to effectively fool multi-object trackers.
• To explore the potential of spatio-temporal motion information for crafting adversarial attacks.
• To evaluate the attack's efficacy and transferability on state-of-the-art multi-object tracking algorithms.

Main Methods:

• Developed a novel Blind-Blur Attack (BBA) leveraging spatio-temporal motion information.
• Employed a perturbation generator trained with a blind-blur loss function.
• Evaluated BBA on TraDeS, CenterTrack, FairMOT, and ByteTrack using MOT-Challenge datasets (MOT16, MOT17, MOT20).

Main Results:

• BBA successfully rendered targets invisible to trackers while causing background elements to be perceived as moving targets.
• The attack drastically reduced the MOTA (Multiple Object Tracking Accuracy) for TraDeS and ByteTrack.
• Demonstrated high transferability of the BBA method across various state-of-the-art tracking algorithms.

Conclusions:

• The proposed Blind-Blur Attack (BBA) is an effective method for degrading multi-object tracker performance.
• Spatio-temporal motion information provides a potent avenue for developing sophisticated adversarial attacks.
• BBA highlights significant vulnerabilities in current deep multi-object tracking systems.