Increasing Neural-Based Pedestrian Detectors' Robustness to Adversarial Patch Attacks Using Anomaly Localization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection is crucial for safety-critical systems like autonomous driving and robotics.
• Adversarial patch attacks pose a significant threat to current object detection systems.
• Existing defenses against these attacks are often insufficient, necessitating new solutions.

Purpose of the Study:

• To develop a robust defense mechanism against adversarial patch attacks in object detection.
• To improve the resilience of neural network systems against manipulated input images.
• To enhance the security and reliability of object detection in real-world applications.

Main Methods:

• A multi-stage method involving image reconstruction using a Deep Convolutional Neural Network.
• Highlighting image discrepancies via a Maximum Error calculation block.
• Localizing anomalous regions using the Isolation Forest algorithm on image fragment histograms.
• Clustering and processing identified anomalous regions for defense evaluation.

Main Results:

• The proposed anomaly localization method demonstrates high resistance to adversarial patch attacks.
• Maintained high-quality object detection performance after applying the defense.
• Achieved a significant improvement in mAP50 from 46.79% to 80.97% for pedestrian detection using YOLOv3 on the INRIAPerson dataset.

Conclusions:

• The developed method effectively defends against adversarial patch attacks.
• The approach shows promise for enhancing the security of object detection systems.
• Anomaly localization is a viable strategy for improving the robustness of AI systems.