Introduction
The Applied Research Centre for Defence and Security (ARC) is a group of data scientists and data science software engineers who focus on the application of cutting-edge technology to problems within the defence and security sector. The centre is staffed by individuals who are full-time employed researchers at the Turing. It enables the UK’s defence and security community to draw on the very best of academia to achieve high impact solutions to the most pressing challenges in the field. The ARC was established in 2019 and has since completed over 20 projects for its funding partners.
Explaining the science
The Applied Research Centre bridges the gap between academia and the defence and security sectors, bringing cutting-edge technology to bear on the unique and challenging problems faced in defence and security. It works on applied problems, primarily in response to operational use cases or policy needs, with a rapid turn-around time.
The emphasis is on delivering usable outputs – such as software code and demonstrators – aligned to practical and targeted research objectives. The ARC runs a flexible operating model, allowing for lightweight tasking and agile retasking as requirements change. They maintain a high level of interactivity with partners through regular meetings, as well as sharing access to GitHub repos and Kanban boards.
Project aims
The ARC allows new technologies and methods to be de-risked for specific applications within defence and security. They produce proof-of-concept demonstrators and provide support on their application to partner data, as well as providing impartial advice on current research and potential applications. This informs partners’ decisions on policy, investment into data science technologies and future-proofing of data storage and processing.
Applications
The ARC provides a lightweight and flexible model for scoping and de-risking the application of data science to defence and security problems.
The team works on short-term applied research projects across the domain of deep learning. Projects are scoped collaboratively with funding partners and these stakeholders are a crucial part of the project’s development, through regular check-ins or direct collaboration. Projects typically deliver a technical report to stakeholders, alongside code, containing experimental results as well as commentary and recommendations.
Model Similarity
With the widespread adoption of deep neural networks across a range of application domains, defending against adversarial attacks is a clear challenge. This project investigated transfer attacks and considered whether we could identify model features or measures of similarity between surrogate and target models that can help identify whether transfer attacks are more likely to succeed. Transfer attacks were simulated between pairs of models to examine the correlation of attack success with model features and model-pair similarity metrics, such as the cosine similarity between loss derivatives.
Voice Liveness Detection
This project sought to demonstrate the applicability of transformer-based models for the defence against audio replay attacks on speaker verification systems. Researchers examined a range of pretrained audio models, with varying architecture and pretraining data, for adaptation to this task and provided experimental results on benchmark datasets.
Selected publications
Low-Cost Model Selection for Transformers (LoCoMoSeT)
Dable-Heath, E., Swatton, P., Roberts, J. and Bishop, J. 2024
Model Similarity Phase 2: Dataset Similarity
Swatton, P., Knight, J. and Bishop, J., 2023