DS/CSS Seminar: David Ifeoluwa Adelani
12:00 pm -
Mitigating Adversarial Attacks on Private User Data and Language Models
A Data Science/Computational Social Science seminar
Machine Learning approaches to Natural Language Processing tasks benefit from a comprehensive collection of real-life user data. At the same time, there is a clear need for protecting the privacy of the users whose data are being collected and prevent the abuse of machine learning models trained on user data. First, I will describe different de-identification methods to protect the leakage of users sensitive information such as credit card number, identity, and patient record in the transcripts of voice interactions systems (like Alexa or Siri) by replacing sensitive parts with benign alternatives, and evaluate the impact of de-identified data on training downstream NLP tasks. Second, I will describe adversarial attacks on neural language models to generate fake reviews for the purpose of influencing the buying decisions of online shoppers. Specifically, I will describe a low-skilled threat model obtained by fine-tuning GPT-2 model for generating sentiment-preserving and fluent online fake reviews which are oftentimes indistinguishable to humans. However, machine-level detection shows some promise.
David Ifeoluwa Adelani is a doctoral student in computer science at Saarland University, Saarbrücken, Germany. His current research focuses on the security and privacy of users’ information in dialog systems and online social interactions. He is also actively involved in the development of natural language processing datasets and tools for low-resource languages, with special focus on African languages.