Security Challenges in Natural Language Processing Models
Abstract:
The recent advancements in large language models, exemplified by ChatGPT and its successors, have significantly expanded our comprehension of the capabilities of AI systems. Despite these remarkable achievements, NLP models are confronting security challenges, particularly in relation to various attacks targeting language models. This seminar seeks to provide an overview of the key security challenges encountered in NLP. It will primarily focus on highlighting research areas that have been advanced by Dr Qiongkai Xu’s recent work, including imitation attacks, backdoor attacks, and data reconstruction attacks. By delving into these topics, the seminar aims to enhance our understanding of the security landscape in NLP and to facilitate discussions on potential mitigation strategies.
Biodata:
Dr. Qiongkai Xu is a lecturer at Macquarie University, having earned his PhD from the Australian National University and previously served as a research fellow at the University of Melbourne. His research primarily focuses on Natural Language Processing, Privacy & Security, Machine Learning and Data Mining. Recently, his attention has been directed towards auditing machine learning models, specifically in two areas: 1) identifying and addressing privacy and security issues in ML/NLP models and their applications and 2) developing comprehensive evaluation theory and methods for ML/NLP models from various perspectives.