Event Date
Speaker: Yue Xing, PhD Candidate, Purdue University
Title: "Statistical Understanding of Adversarial Training"
Abstract: Artificial intelligence plays an important role in various disciplines, such as auto-driving and information technology. Recent studies reveal that these dedicated models are vulnerable to adversarial attack, i.e., the predicting label may be changed even if the testing input has an unaware perturbation. Most existing studies to defend such attacks aim to develop computationally efficient adversarial learning algorithms, but without a thorough understanding of the statistical properties of these algorithms.
We provide theoretical understandings of adversarial training to figure out potential improvements. First, we study the algorithmic stability of adversarial training. We reveal that the stability of the vanilla adversarial training is sub-optimal, and a simple noise injection method can improve the stability. Second, we focus on how artificially generated data improves adversarial training. It is observed that utilizing synthetic data improves adversarial robustness. We use a theory to explain the reason behind this observation.
SEMINAR TIME/DATE: Friday, January 27, 11:00am
LOCATION: MSB 1147