Security in Transformer Visual Trackers: A Case Study on the Adversarial Robustness of Two Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Visual object tracking is crucial for autonomous driving systems.
• Transformer models, utilizing self-attention, are increasingly used in visual tracking.
• The security vulnerabilities of transformer models in this domain are not well understood.

Purpose of the Study:

• To investigate the security of transformer-based visual tracking against adversarial attacks.
• To evaluate the impact of adversarial perturbations on tracking performance.
• To quantify the effectiveness of different attack strategies.

Main Methods:

• Generated adversarial examples on video sequences, considering frame-by-frame temporal motion.
• Implemented adversarial attacks to degrade transformer-based visual tracking performance.
• Conducted experiments on benchmark datasets (OTB100, VOT2018, GOT-10k).

Main Results:

• Adversarial examples effectively reduced the performance of transformer-based visual trackers.
• White-box attacks demonstrated the highest effectiveness, exceeding 90% attack success rates.
• The study confirms the vulnerability of these models to security threats.

Conclusions:

• Transformer-based visual tracking systems are susceptible to adversarial attacks.
• Security assessments are essential for safety-critical applications like autonomous driving.
• Further research is needed to develop robust defense mechanisms against such threats.