An adversarial example attack method based on predicted bounding box adaptive deformation in optical remote sensing...

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Existing adversarial attacks are ineffective for real-time optical remote sensing object detectors (YOLO series).
• Current methods struggle with optical remote sensing images due to unsuitable adversarial perturbation mechanisms.
• Improving adversarial robustness of single-stage detectors remains a challenge.

Purpose of the Study:

• To propose a novel adaptive deformation method (ADM) for fooling YOLO object detectors.
• To enhance adversarial attacks against YOLOv4 and YOLOv5 in optical remote sensing.
• To provide a more effective scheme for evaluating adversarial resilience.

Main Methods:

• Introduced an Adaptive Deformation Method (ADM) to generate adversarial perturbations.
• Developed Adaptive Deformation Method Iterative Fast Gradient Sign Method (ADM-I-FGSM) and Adaptive Deformation Mechanism Projected Gradient Descent (ADM-PGD).
• Utilized the length-to-width ratio of prediction boxes to determine deformation trends for perturbations.

Main Results:

• The proposed ADM achieved a higher adversarial success rate compared to state-of-the-art methods.
• ADM-based attacks effectively fool YOLOv4 and YOLOv5 detectors.
• The generated adversarial perturbations demonstrated a superior adversarial effect.

Conclusions:

• The adaptive deformation method offers a more effective approach for adversarial attacks on YOLO detectors in optical remote sensing.
• This attack scheme provides a valuable tool for assessing the adversarial resilience of these models.
• The findings highlight the need for robust defense mechanisms against sophisticated adversarial attacks.