SSIM-Based Autoencoder Modeling to Defeat Adversarial Patch Attacks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Object detection systems, crucial for autonomous vehicles, utilize deep learning for real-time processing on edge devices.
• Edge devices are susceptible to physical adversarial attacks, posing security risks.
• Existing object detection models like YOLO and Faster-RCNN are vulnerable to such attacks.

Purpose of the Study:

• To investigate the effectiveness of adversarial patch attacks (Adv-Patch, Dpatch) on traffic sign detection in autonomous vehicles.
• To propose and evaluate a defense mechanism against these adversarial attacks.

Main Methods:

• Implemented traffic sign detection using You Only Look Once (YOLO) and Faster-RCNN.
• Simulated adversarial patch attacks using Adv-Patch and Dpatch to cause misdetections.
• Developed an image reconstruction defense method utilizing autoencoders and Structural Similarity Index Measure (SSIM).

Main Results:

• Adversarial patch attacks were confirmed to succeed with high probability in causing misdetections of traffic stop signs.
• The proposed autoencoder and SSIM-based reconstruction method demonstrated effective defense capabilities.
• The defense method achieved a mean Average Precision (mAP) of 91.46% even under dual adversarial attacks.

Conclusions:

• Deep learning-based object detection for autonomous vehicles is vulnerable to physical adversarial attacks.
• The proposed image reconstruction method offers a robust defense against adversarial patch attacks.
• The research contributes to enhancing the security and reliability of autonomous driving systems.