Detection of AI-Generated Images Using Combined Uncertainty Measures and Particle Swarm Optimised Rejection Mechanism

Distinguishing AI-generated images from genuine photographs is becoming increasingly difficult as generative models produce images that are ever more photorealistic. This talk presents a detection framework for distinguishing AI-generated images from real images using multiple uncertainty signals. It combines Fisher Information, Monte Carlo Dropout entropy, and Gaussian Process predictive variance to evaluate prediction reliability. Particle Swarm Optimisation learns optimal weights and rejection thresholds. Tested across generators such as Stable Diffusion, Midjourney, and StyleGAN3, the combined uncertainty approach maintains strong robustness under distribution shifts and adversarial attacks, rejecting most misclassified synthetic images while preserving accurate predictions for natural images. Published paper

Speaker: Dr Rahul Yumlembam (Research Fellow), Northumbria University, Newcastle upon Tyne, UK

NOTE: From 11:00-13:00, those in the North East England (especially from the five NE universities – Newcastle, Northumbria, Durham, Sunderland and Teesside and any NE industry contacts) will be welcome to stay back for an hour of introductions and, if you are attending at Northumbria (venue), please stay back for lunch at the venue, after the introductions. Thank you.

Registrant details will be shared with event organisers in Northumbria University

This is a hybrid event (in person and online). If attending remotely, please join using the Teams link below:
https://teams.microsoft.com/meet/38616455579988?p=3CpAErPCcNAYM82K09
Meeting ID: 386 164 555 799 88
Passcode: y3QR7Nu7