Date and time: Tuesday 23 June 2026, 14:30-15:30 CEST
Speaker: Mark Dras, School of Computing at Macquarie University, Australia
Title: Applications of Differential Privacy to Text
Where: Harry Nyquist, Malvinas väg 10 (only in person attendance)
Host: Tobias Oechtering oech@kth.se
Bio: Mark Dras is a professor at the School of Computing, Macquarie University in Australia. He has spent most of his career there, after a postdoctoral fellowship with the Institute for Research in Cognitive Science at the University of Pennsylvania. Dras works on machine learning and artificial intelligence, with a particular focus on Natural Language Processing and Large Language Models, especially on topics of privacy and security. His funding for work in these areas has come from the Australian Research Council, the Australian Department of Defence, Oracle, the Medical Research Future Fund, Office of National Intelligence, Cooperative Research Centres, and elsewhere.
Mark is currently Treasurer of the Asia-Pacific Chapter of the Association for Computational Linguistics, the premier international scientific and professional society for people working on computational treatment of human language. He has also been appointed to the Australian Research Council’s College of Experts (2025-2027).
Abstract: Differential Privacy (DP) was originally conceived of as applying to tabular-style data; a bit over a decade ago, it was also applied to machine learning, for the most part in the domain of images. This talk will look at applying DP to language models, and consider the questions: What does it mean to make text private? and What issues are raised for DP by applying it to the text domain? Specifically, the will cover three pieces of work in the domain of text: (1) The original work of ours that proposed to use metric DP to obfuscate text, with the goal of preventing leakage of sensitive information. (2) An application of metric DP in the training of language models using directional privacy via the von Mises-Fisher distribution, and the issues that raises regarding how to empirically calibrate DP variants across frameworks. (3) Some current work on calibrating local DP as is increasingly widely used in privatised text rewriting, via the use of distinguishability auditing techniques.
