Evaluating AI defences
Test beyond the examples a defender has seen.
How can we tell whether an AI defence will handle a new generator, a changed attack or a different setting?
A defence needs evaluation that extends beyond familiar examples. This exploratory area considers how generative methods can create controlled fraud, impersonation and synthetic-media scenarios for training and testing. The research challenge is to keep evaluation independent, measure generalisation and explain the limits of a result, including uncertainty and the cost of intervention.
Research areas
Create controlled scenario variations with explicit labels and assumptions, then check their relevance to the intended defensive task.
Separate training and evaluation by generator, campaign or time to examine how results transfer to unfamiliar cases.
Measure detection quality alongside false alarms, response latency, computational cost and human review effort.
Compare content-only decisions with combined behavioural, contextual and provenance evidence.
Evaluate when a system should request further verification or defer a decision to a person.
Publications & related resources
Projects, publications and research materials will be added here.
Contribute to this research.
Explore student research, internships and collaboration with the lab.