A Survey and Evaluation of Adversarial Attacks in Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence Security

Background:

• Deep learning models excel in computer vision but are vulnerable to adversarial examples.
• These vulnerabilities pose risks in critical applications like autonomous driving and surveillance.
• Existing research on adversarial attacks primarily focuses on image classification, with limited analysis on object detection systems.

Purpose of the Study:

• To introduce a novel taxonomic framework for adversarial attacks on object detection.
• To synthesize existing robustness metrics for object detection.
• To empirically evaluate state-of-the-art attack methodologies on various object detection models.

Main Methods:

• Developed a new taxonomy for categorizing adversarial attacks specific to object detection.
• Synthesized and applied existing robustness metrics.
• Conducted empirical evaluations of leading attack methods on traditional and vision-language pre-trained object detectors using open-source implementations.

Main Results:

• Provided a comprehensive analysis of adversarial attack characteristics across diverse object detection architectures.
• Identified key insights into the effectiveness of different attack strategies.
• Highlighted significant research gaps and emerging challenges in the field.

Conclusions:

• Established a foundational understanding for developing more robust object detection models.
• Emphasized the critical need for standardized evaluation protocols in adversarial robustness research.
• Underscored the ongoing threat of adversarial examples to AI systems and the necessity for enhanced security measures.