A team from Keele University led by doctoral student Saman Magsouda has developed an artificial intelligence model, FDIP-NET, that recognizes fake face images by analyzing color and the nature of brightness variation across a photo. The work was published in the journal Frontiers in Imaging.
Synthetic photos and videos of faces, known as "deepfakes," have become easier to create with the development of generative AI. Therefore, the ability to reliably distinguish genuine images from fake ones is important for trust in digital media and for combating identity theft.
The FDIP-NET model combines ordinary color information with three additional measures that describe how brightness changes in different parts of an image. These measures are inspired by mathematical ideas from fluid dynamics, but they do not model actual fluid motion - they merely characterize the local structure of the picture. In addition, the model estimates confidence in its predictions and studies how detection errors relate to image properties: blur, brightness and contrast.
The researchers tested the model on the publicly available face image sets DFRI, DF-wild and RFF-140K. The average accuracy on the tests was 98.654%, 98.952% and 99.669%, respectively. These results are averaged over three training runs, with the models trained and tested separately in each set. In an additional check using the official FaceForensics++ c23 split into training, validation and test sets, the model showed an average accuracy of 91% on individual video frames.
The authors emphasize: the advantages of the approach varied depending on the dataset and metric, and individual tests showed that reliable detection of forgeries on unfamiliar datasets remains a task for future research. In other words, high accuracy within familiar data does not guarantee equally stable performance in new conditions.
Lead author Saman Magsouda noted: "As AI-generated faces become increasingly convincing, assessing the authenticity of images is becoming ever more important for trust in digital media and for combating identity fraud and disinformation. Our results show how brightness patterns can complement conventional features extracted by AI, and why prediction confidence and detection errors deserve attention alongside accuracy. This creates a foundation for further research into detectors that will work more reliably on unfamiliar images and in real-world conditions."