Speaker

Thai Le

Abstract

Responsible AI models should not only be accurate but also resilient to security threats and transparent to users. Despite their high performance, many recent Natural Language Processing models are vulnerable to adversarial text attacks. As these models become more complex, understanding their decision-making process and ensuring compliance with societal standards is crucial. To illuminate these topics, this talk first introduces the first and most comprehensive collection of over 2.5 million human-written text perturbations available online. It explores how this collection can expose the adversarial vulnerabilities of existing NLP algorithms, allowing malicious individuals to change their predictions by minimally manipulating the input texts, particularly those designed to deal with sensitive topics like COVID-19 vaccines and presidential campaigns on social networks. The talk further examines how these vulnerabilities can enable manipulation of not only predictions but also their explanations. It will then discuss the societal consequences of such manipulations on critical AI applications, especially in medical, health care, and legal domain.

Bio

Dr. Thai Le is currently an Assistant Professor at the University of Mississippi. He got his doctorate degree from the College of Information Science and Technology (IST) at the Pennsylvania State University. He was awarded the IST PhD Student Award for Research Excellence and was a DAAD Postdoctoral Fellow. He has industry research experience at Amazon Alexa, Yahoo Research and VMWare. Dr. Le’s mission is to enhance the robustness, safety, and explainability of Machine Learning and Artificial Intelligence, especially in the Natural Language Processing domain, ensuring that the society can harness their power with confidence and clarity. His work has been published in venues such as ACL, EMNLP, NAACL, AAAI, AAMAS, KDD, WebConf and CHI, and has been featured in ScienceDaily, DefenseOne, and Engineering and Technology Magazine. His profile can be viewed at lethaiq.github.io.