Building a Universal Detector of AI-Generated Images Without Training on Them

Deep generative models are now capable of generating synthetic images with very high visual realism, often indistinguishable from real-world photographs. Such AI-generated images (AIGIs) can pose serious security concerns if used maliciously. Conventional AIGI detection methods are based on supervised learning and may have limited generalization ability. In this paper, we build a novel universal detector of AIGIs without the need to perform training on these images. Starting with a study on the effectiveness of various pretrained image models for the AIGI detection task, we then chose to build our detector based on the features of the popular CLIP model. Unlike existing methods, we use a small number of real images and their carefully processed counterparts as AIGI proxies during training, combined with a novel margin-based loss to promote generalization. Extensive experiments demonstrate the effectiveness of our method, outperforming existing supervised methods while not using any AIGI for training.