Overview
Explore a groundbreaking study on exploiting correlation biases in major computer vision systems to craft adversarial images. Learn how objects commonly found together in nature create strong correlations that lead to detection biases in AI models. Discover examples like how round shapes near dogs are often misclassified as frisbees, while weakly correlated objects like stop signs and pizza become harder to detect together. Examine the researchers' methods for generating adversarial images using popular object detection models like RetinaNet, YOLOv3, and TinyYOLOv3 trained on the COCO dataset. Gain insights into the implications of these biases for computer vision applications and potential mitigation strategies in this 27-minute Black Hat conference talk by Masaki Kamizono, Yin Minn Pa Pa, and Paul Ziegler.
Syllabus
Hiding Objects from Computer Vision by Exploiting Correlation Biases
Taught by
Black Hat