Foundations for secure deep learning

Lead Research Organisation: University of Oxford
Department Name: Computer Science


In this research project I explore foundations for safe and secure deep learning including:
-Systems capable of detecting anomalies/out-of-distribution behaviour using Bayesian deep learning methods and ensembles.
-Differentially private deep learning systems for learning over extended periods of time or across related simultaneous contexts with managed privacy leakage.
-Deep learning architectures optimized for secure multi-party computation.

This research project is linked to the EPSRC Cyber Security Research Theme

Planned Impact

It is part of the nature of Cyber Security - and a key reason for the urgency in developing new research approaches - that it now is a concern of every section of society, and so the successful CDT will have a very broad impact indeed. We will ensure impact for:

* The IT industry; vendors of hardware and software, and within this the IT Security industry;

* High value/high assurance sectors such as banking, bio-medical domains, and critical infrastructure, and more generally the CISO community across many industries;

* The mobile systems community, mobile service providers, handset and platform manufacturers, those developing the technologies of the internet of things, and smart cities;

* Defence sector, MoD/DSTL in particular, defence contractors, and the intelligence community;

* The public sector more generally, in its own activities and in increasingly important electronic engagement with the citizen;

* The not-for-profit sector, education, charities, and NGOs - many of whom work in highly contended contexts, but do not always have access to high-grade cyber defensive skills.

Impact in each of these will be achieved in fresh elaborations of threat and risk models; by developing new fundamental design approaches; through new methods of evaluation, incorporating usability criteria, privacy, and other societal concerns; and by developing prototype and proof-of-concept solutions exhibiting these characteristics. These impacts will retain focus through the way that the educational and research programme is structured - so that the academic and theoretical components are directed towards practical and anticipated problems motivated by the sectors listed here.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/P00881X/1 01/10/2016 31/03/2023
1938170 Studentship EP/P00881X/1 02/10/2017 30/09/2021 Sebastian Farquhar