Information Sciences Seminar——Causality-based Neural Network Repair
报告人：Jun Sun (Singapore Management University)
时间：2022-03-02 14:30-15:30
地点：腾讯会议：397-978-504
Abstract:
Neural networks have had discernible achievements in a wide range of applications. The wide-spread adoption also raises the concern of their dependability and reliability. Similar to traditional decision making programs, neural networks can have defects that need to be repaired. The defects may cause unsafe behaviors, raise security concerns or unjust societal impacts. In this work, we address the problem of repairing a neural network for desirable properties such as fairness and the absence of backdoor. The goal is to construct a neural network that satisfies the property by (minimally) adjusting the given neural network’s parameters (i.e., weights). Specifically, we propose CARE (CAusality-based REpair), a causality-based neural network repair technique that 1) performs causality-based fault localization to identify the ‘guilty’ neurons and 2) optimizes the parameters of the identified neurons to reduce the misbehavior. We have empirically evaluated CARE on various tasks such as backdoor removal, neural network repair for fairness and safety properties. Our experiment results show that CARE is able to repair all neural networks efficiently and effectively. For fairness repair tasks, CARE successfully improves fairness by 61.91% on average. For backdoor removal tasks, CARE reduces the attack success rate from over 98% to less than 1%. For safety property repair tasks, CARE reduces the property violation rate to less than 1%. Results also show that thanks to the causality-based fault localization, CARE’s repair focuses on the misbehavior and preserves the accuracy of the neural networks.
Bio:
SUN, Jun is currently a professor at Singapore Management University. He received Bachelor and PhD degrees in computing science from National University of Singapore (NUS) in 2002 and 2006. In 2007, he received the prestigious LEE KUAN YEW postdoctoral fellowship. He has been a faculty member of SUTD since 2010 and was a visiting scholar at MIT from 2011-2012. Jun's research interests include software engineering, cyber-security and formal methods. He is the co-founder of the PAT model checker. He has published more than 250 articles and conference papers including top conferences in multiple areas.
