Adaptable AI in the Wild for Defence and Security

Research Theme

Sensor Signal Processing

Aim

To develop a defence-ready, low-cost alternative to AI post-training that lets air-gapped AI systems rapidly adapt in the wild 

Objectives

1. Build a local “module bank” of lightweight adapters (LoRAs) covering defence relevant sensors and tasks.
2. Train a learned router that picks and combines the right modules for new scenes, using crossmodal alignment and simple ondevice self-supervision
3. Prove the system on air‑gapped edge hardware, measuring speed, accuracy, and reliability under distribution shift

Description

We will develop a low-cost alternative to AI post-training that relies on exploiting a local bank of neural modules, implemented with parameter-efficient adapters (e.g., LoRA) and related lightweight components, for deployment on defence and security platforms. This will allow an AI in the wild—operating in air‑gapped or contested settings—when exposed to a new scenario (e.g. an unidentified object, a change of scene) to rapidly adapt to what it senses using the most appropriate modules, guided by a learned router that selects and composes modules and supports lightweight on‑device self‑supervised retraining to track distribution shift. We will optimise this module bank using both generative and discriminative tasks across defence‑relevant modalities (e.g. natural images, EO/IR, SAR/sonar, RF, audio) and develop algorithms that are able to express a new task/modality in terms of these tasks/modalities to identify the most appropriate modules via cross‑modal alignment into a shared task/sensor embedding space.

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. Further information on English language requirements for EU/Overseas applicants.

Full funding is available for this position