Vulnerability Relationship between Feature Vector Scale and Convolutional Neural Networks with Adversarial Examples
Vulnerability Relationship between Feature Vector Scale and Convolutional Neural Networks with Adversarial Examples
October 21, 2021
Guest Speaker: Dr. Sang-Woong Lee, Professor, Gachon University
Moderated by Dr. Reza Rawassizadeh, Assistant Professor of Computer Science
Abstract: In the field of image classification, Deep Convolutional Neural Networks can misclassify images by perturbation noise. Images that are artificially created to cause misclassification are called adversarial examples. There are various conjectures about why DCNN is vulnerable to noise such as adversarial examples. We hypothesize that DCNN is vulnerable to noise because of ‘unfair data learning’. Furthermore, we assume that ‘unfair data learning’ learns the scale of feature vectors differently in feature spaces. Thus, the trained data will exhibit different vulnerabilities against noise depending on the scale of the feature vector. We use DCNN and CIFAR-10 datasets to conduct vulnerability tests for each scale section of feature vectors. Vulnerability experiments compare cosine similarity from each feature vector of the original and noisy images and observe the error rate by scale sections. Experimental results showed sensitive results with low cosine similarity and a high error rate in the small-scale section. On the other hand, in the high-scale section, the result shows robustness for noise with high cosine similarity and low error rate.
Speaker Bio: Sang-Woong Lee received the B.S. degree in electronics and computer engineering and the M.S. and Ph.D. degrees in computer science and engineering from Korea University, Seoul, South Korea, in 1996, 2001, and 2006, respectively. From June 2006 to May 2007, he was a Visiting Scholar with the Robotics Institute, Carnegie Mellon University. From September 2007 to February 2017, he was a Professor with the Department of Computer Engineering, Chosun University, Gwangju, South Korea. He is currently a Professor with the Department of Software, Gachon University. His current research interests include face recognition, computational aesthetics, machine learning, and medical imaging analysis.