Keele researchers develop physics inspired deepfake detection tool
Researchers at Keele University have developed a physics-inspired AI model that achieved high accuracy in tests distinguishing genuine facial images from AI-generated or manipulated images.
The research team, led by PhD candidate Saman Maghsoudi from Keele's School of Computer Science and Mathematics, have created a tool which analyses colour information and patterns of brightness within an image to help distinguish genuine facial images from AI-generated or manipulated ones.
Creating synthetic photographs or videos of faces, known as "deepfakes", has become even easier in the age of generative AI, meaning the ability to reliably spot genuine or fake images is becoming increasingly important for trust, security, and authenticity in things like digital media and combatting identity theft.
The model proposed by the research team, including Dr Sangeeta Sangeeta and Dr Nadia Kanwal, is known as FDIP-NET. It combines conventional colour-image information with three additional measurements derived from how brightness varies across an image. These measurements are inspired by mathematical concepts from fluid dynamics and describe local image structure without modelling physical fluid motion.
The model analyses colour and spatial variations in image brightness, rather than tracking a person’s movement or changes over time. The framework also calibrates prediction confidence and examines how detection errors are associated with image properties such as blur, brightness and contrast. The researchers evaluated it using publicly available facial-image datasets known as DFRI, DF-wild and RFF-140K.
In their experiments, FDIP-NET achieved mean test accuracies of 98.654%, 98.952% and 99.669% on DFRI, DF-wild and RFF-140K respectively. These results were averaged across three training runs, with models trained and tested separately within each dataset. In an additional evaluation using the official FaceForensics++ c23 training, validation and test splits, the model achieved 91% mean accuracy on individual video frames.
The findings highlight the potential of combining conventional colour-image features with structural information derived from image brightness to support deepfake detection. The benefits varied across datasets and evaluation measures, and separate tests showed that reliable detection on unfamiliar datasets remains a challenge for future research. The study was published in Frontiers in Imaging.
Lead author Saman Maghsoudi said: “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. Our findings show how patterns of brightness within images can complement conventional AI features, while highlighting why prediction confidence and detection errors deserve attention alongside accuracy. This provides a foundation for further research towards detectors that work more reliably on unfamiliar images and under real-world conditions.”
