|제목||[외부초청세미나] Microsoft의 나태식 박사(7월 30일(화), 오후 2시, 장소 : 반도체관 400112호)|
1. Title : Noise Robust and Secure Machine Vision Techniques
2. Speaker : 나태식박사
3. Time : 7월 30일(화), 오후 14:00
4. Place : 반도체관 400112호
Deep learning based machine vision systems have shown to be vulnerable to small input perturbations generated by adversaries. And image distortions due to energy-quality trade off on resource constrained devices degrade target accuracy of an image classifier and object detection network. Meanwhile, critical applications like autonomous driving always require high fidelity results of a deep learning based machine vision system even under adversarial and noisy environments.
To address this challenge, I would like to talk about how to improve robustness of a machine vision system efficiently against adversarial attacks and general image distortions. First, I will talk about cascade adversarial machine learning and additional regularization using a unified embedding for image classification and low level (pixel level) similarity learning.
Next, I will introduce simple yet efficient mixture of pre-processing experts model for noise robust object detection network.
Taesik Na joined Microsoft's AI and Advanced Architecture group as a hardware engineer in 2018. He earned his Ph.D. degree from Georgia Tech in 2018 under the supervision of Prof. Saibal Mukhopadhyay, and received MS degree from Seoul National University early 2008 under the supervision of Prof. Deog-Kyoon Jeong. From 2008 to 2013, he was a circuit design Engineer at Samsung. His current research interests include secure and robust deep learning and real-time machine learning inference.
He was a recipient of the Korea Foundation for Advanced Studies (KFAS) Scholarship and the Kwanjeong Educational Foundation Doctoral Fellowship during 2006–2008 and 2013–2018, respectively. He won the first place in the research track at the Institute for Information Security & Privacy’s Cybersecurity Demo Day Finale in 2018.