Keele deepfake detector scores highly in tests, but its makers say it can't yet be trusted on unfamiliar images
A team at Keele University has built an AI model, FDIP-NET, that spots fake faces by looking at how brightness changes across a picture. It scored above 98% on three test sets, but its published paper shows performance fell sharply on images unlike those it was trained on.

Researchers at Keele University, just up the road from Newcastle-under-Lyme, have built an AI model that tries to tell real photographs of faces from AI-generated or doctored ones by looking at how brightness changes across the picture.
The model is called FDIP-NET. It is described in a peer-reviewed paper published in the journal Frontiers in Imaging on 30 October, and in a Keele University press release on 13 November. The lead author is PhD candidate Saman Maghsoudi, working with Dr Sangeeta Sangeeta and Dr Nadia Kanwal. The press release places the team in the School of Computer Science and Mathematics; the paper gives the School of Computing and Mathematics.
As AI-generated faces become more convincing, assessing image authenticity is increasingly important for trust in digital media and efforts to tackle impersonation and misinformation.Saman Maghsoudi, PhD candidate, Keele University
How it works
Most image detectors learn from the colours and shapes in a picture. FDIP-NET feeds in those normal colour channels plus three extra maps worked out from luminance, which is the brightness of each part of the image. The paper calls them a divergence-like response, a speed-like measure based on how sharply brightness changes, and a kinetic-energy-like response.
Advert spaceWant your ad here?300 × 250 boxEnquire about this space →Those names are borrowed from fluid dynamics, the maths of how liquids and gases move. Both the paper and the university are clear that the model doesn't simulate anything flowing. The equations are just used as a way to describe the fine structure of brightness in a face.
The model is built on a ResNet-style network, a common design for image recognition. It also adjusts its confidence scores so they better match how often it is actually right, and the team looked at how its mistakes relate to blur, brightness and contrast in the pictures.
The results
The team tested FDIP-NET on three public collections of real and fake faces, all downloaded from Kaggle. Averaged over three training runs, it scored 98.654% on DFRI, 98.952% on DF-wild and 99.669% on RFF-140K. On individual frames from the widely used FaceForensics++ video benchmark, it averaged 91%.
Those scores come from training and testing the model within each dataset. In other words, the test pictures were new to the model but came from the same collection, made in the same way, as the pictures it learned from.
The catch
The paper is open about the limits. When the team tried the model on datasets it had never seen, it reports "substantial performance degradation". It says that "Strong within-dataset performance does not establish reliable transfer to unseen dataset distributions", and that performance on these datasets does not, on its own, show the model is reliable for identity checks. In those cross-dataset tests, a plain colour-only version of the model did better on average than the version with the extra brightness maps.
The authors also note that their image-level split didn't confirm that the same person or video never appeared in both training and test sets, which can flatter results. They list cross-dataset robustness as unresolved and a job for future work. The press release says the gains varied between datasets and measures, and that reliable detection on unfamiliar data remains a challenge.
The paper says the work received no specific funding, and the authors declare no commercial relationships. It also states that ChatGPT was used to help with clarity and wording, and that the authors verified all scientific content. The datasets are public, but the paper doesn't say whether the model's code will be released, and the university hasn't said either.
Why detection matters
Deepfakes are already common. Ofcom research from June 2024 found that 43% of people aged 16 and over said they had seen at least one deepfake online in the previous 6 months, and only 9% were confident they could spot one. Ofcom's list of harms includes deepfakes that defraud, such as fake adverts and romance scams that misuse someone's identity.
Ofcom describes detection as one of several approaches, alongside prevention, embedding and enforcement, and says there is "no silver bullet". The Keele paper's own results point the same way: a detector that does well on familiar images can struggle with new kinds of fake.
What this means for your business
- There is nothing here to buy or install: FDIP-NET is published research, and its makers say it isn't yet reliable on unfamiliar images.
- If your business uses photos or video for ID checks, signing up customers or approving payments, don't treat an image on its own as proof. Confirm anything important another way, for example by ringing the person on a number you already hold.
- If a supplier sells you a deepfake detector, ask how it performs on images from outside its training data, not just its headline accuracy.
Sources
- Keele University: Keele researchers develop physics inspired deepfake detection tool (13 Nov 2026) (press release) (archived copy)
- Frontiers in Imaging: FDIP-NET: a physics-inspired deepfake image detector leveraging FDIP luminance dynamics and calibrated CNNs (30 Oct 2026) (published research) (archived copy)
- Ofcom: A deep dive into deepfakes that demean, defraud and disinform (23 Jul 2024) (regulator research) (archived copy)
Spotted a mistake? Read our corrections policy and tell us. How this story was put together: the newsroom.