Detection of AI-Generated Sportsman Images Using Deep Learning

AI image generators can now create very realistic photos of sportspersons. Such fake photos are hard to tell apart from real ones and can spread false news in sports media. This paper presents a system that detects whether a sportsman image is a real photograph or AI-generated. We built a new dataset of 5,422 real sports photos from 35 sports and 10,922 AI-generated images made with four generators: SDXL-Turbo, SD-Turbo, Stable Diffusion 1.5 and PixArt-Σ. All images were given the same size and file format, so the model cannot use these as shortcuts. We first trained a ResNet-18 model, which reached 99.82% test accuracy but did not work well on images from new commercial generators. Our final system uses a pretrained CLIP ViT-L/14 image encoder, kept frozen, with a logistic regression classifier. It reaches 99.57% test accuracy and a ROC-AUC of 0.9999, and five-fold cross-validation gives 99.60% ± 0.16%. It also detects all PixArt-Σ test images even when PixArt-Σ is not used in training. With a decision threshold of 0.70, it detects 100% of the AI-generated test images. The system runs as a web application and takes about 27 ms per image on a T4 GPU. Key Words: AI-generated image detection, deepfake detection, CLIP, transfer learning, diffusion models, sports media.