Vote for this team on Twitter before April 12 at 6 p.m., using the hashtag #cyberfinaleteam7. Votes count toward the $2,000 "People's Choice" award.
"Deep Security: Toward Robust Deep Learning"
Taesik Na and Jong Hwan Ko
School of Electrical & Computer Engineering
A successful deep learning-based computer vision task is perceived as a key enabler for autonomous vehicles. However, there have been numerous reports that deep-learning classifiers are vulnerable to small input perturbations that have been carefully generated by adversaries. Vulnerabilities in deep learning can become potential threats to successful autonomous driving. The objective of this research is to build robust deep-learning classifiers for various adversarial attacks in order to better protect self-driving cars. To address this challenge, we propose embedding space for both classification and low-level (pixel-level) similarity learning that will ignore unknown pixel level perturbation. We also propose cascade adversarial training, which transfers the knowledge of the end results of adversarial training. This proposed approach shows improved accuracy compared to the current state-of-the-art adversarial training and ensemble adversarial training methodologies.
About the Students