The Cost of Learning from the Best - How Prior Knowledge Weakens the Security of Deep Neural Networks

Explore the security implications of using pre-trained image classification models in object detection systems through this 30-minute Black Hat conference talk. Discover alarming findings from a measurement study revealing that mainstream object detection models commonly incorporate winning ImageNet contest models as initial layers for low-level feature extraction. Examine how this practice of leveraging prior knowledge may inadvertently compromise the security of deep neural networks. Gain insights from speakers Yunhan Jia, Zhenyu Zhong, Yulong Zhang, Qian Feng, Tao Wei, and Yantao Lu as they present their research and discuss potential vulnerabilities in popular AI vision systems.