Keele researchers build a deepfake detector that reads patterns of light in faces
The model, FDIP-NET, scored above 98% on three public test sets, though its makers say spotting fakes in unfamiliar images is still an open problem.
Researchers at Keele University have built an AI model that tells real photographs of faces from AI-generated or doctored ones by looking at how brightness changes across the image.
The model, called FDIP-NET, was developed in Keele's School of Computer Science and Mathematics and published in the journal Frontiers in Imaging. The lead author is PhD candidate Saman Maghsoudi, working with Dr Sangeeta Sangeeta and Dr Nadia Kanwal.
How it works
Most detectors look only at the colours and shapes a standard image model picks up. FDIP-NET adds three extra measurements of how brightness varies from one small area of a picture to the next. The maths behind them is borrowed from fluid dynamics, but the university is clear that the model doesn't simulate anything flowing; it uses the equations to describe the local structure of the image.
Advert spaceWant your ad here?300 × 250 boxEnquire about this space →The model also reports how confident it is in each answer, and the team looked at how its mistakes relate to blur, brightness and contrast.
Our findings show how patterns of brightness within images can complement conventional AI features.Saman Maghsoudi, Keele University
The results, and the catch
Tested on three publicly available sets of real and fake faces, FDIP-NET averaged 98.7%, 99.0% and 99.7% accuracy. On individual frames taken from videos in the widely used FaceForensics++ benchmark it averaged 91%.
Those figures come from training and testing the model separately within each dataset. The researchers say the gains varied between datasets and measures, and that detecting fakes reliably in images unlike the ones a model was trained on is still an open challenge. That's the situation a detector faces in the real world, and it's where they plan to take the work next.
The university hasn't said whether the tool or its code will be released.
What this means for your business
There's nothing here you can buy or install yet: it's research, and its makers say spotting fakes in unfamiliar images is still unsolved. If your business relies on photos for ID checks or for approving payments, don't treat a picture on its own as proof. Confirm anything important another way, for example by ringing the person on a number you already have.
Sources
Spotted a mistake? Read our corrections policy and tell us. How this story was put together: the newsroom.