I recently undertook a six-week academic visit to the UK with the support of funds donated from the Department of Mathematics and Statistics at the University of Melbourne, the Belz Fund and OPTIMA. This trip was an extension of my research which I am currently undertaking in collaboration with the Australian Bureau of Statistics as part of an industry PhD project through OPTIMA.

My visit began and ended at the University of Manchester, where I was hosted by Professor Mark Elliott. Mark is a world leader in privacy and confidentiality, having been a key researcher in the field since the mid-1990s. I sought out Mark’s expertise as it aligns with my own research regarding privacy-preserving publication of sensitive information. The key challenge in this field is to balance the competing goals of data utility and data privacy.

During my stay I was asked to give a talk to the Social Statistics Department of the university. This talk, titled “Generating Private Databases with Markov-Chain Monte Carlo (MCMC)”, explained my research relating to data synthesis, a topic which has been a major focus of Mark’s work in recent years. Data synthesis relates to generating a private or “synthetic” version of the database which can be published without fear of revealing the sensitive information of the individuals contained in the original database. The research that I am currently undertaking with the ABS and my supervisors, Professor Kate Smith-Miles and Professor Peter Taylor, generates these databases by way of an MCMC algorithm – crawling around the space of synthetic databases before locating one that sufficiently balances privacy and utility.

Following my initial stay in Manchester I travelled south to Oxford University to begin my academic visit with Professor Sam Cohen. Sam is part of the Mathematical and Computational Finance and Data Science Groups in Oxford, working in stochastic analysis and mathematical finance. As he noted to me, Sam has spent “the majority of his career working on high-dimensional probability.” Together we worked on a very general method to modify the training regime of machine learning algorithms to adhere to Differential Privacy, a formal definition of privacy. This has become an increasingly important problem as concerns over the use of personal information in training deep learning models have grown with their development in recent years. Our proposed method treats the learning process of an algorithm as a Markov chain, crawling around the space of parameters in a similar manner to the algorithm that I have been developing with the ABS. The key mathematical challenge to overcome here is lifting the algorithm dynamics to a very high-dimensional space and performing the subsequent analysis.

In addition to these visits, I ventured further south to speak to the Computational Privacy Group at Imperial College, London, where I presented a modified version of the talk I gave in Manchester. I also met with Luc Rocher and Alex Edmonds back in Oxford and Professor Graham Cormode at the University of Warwick. Finally, I returned to Manchester to attend and present at a workshop organised by Mark on “Innovations in Data Synthesis,” before embarking on the long journey home.

I would like to sincerely thank the Department of Mathematics and Statistics at the University of Melbourne, the Belz Fund and OPTIMA for giving me the opportunity to go on this trip and engage with experts in my field.