Intermediate representations are powerful AI-generated image detectors

The researchers propose a search-based AI-generated image detection method that exploits the sensitivity of intermediate-layer data embeddings, outperforming training-free and training-based SOTA methods on GenImage and Forensics Small. The method compares the similarity between the original image embedding and the perturbed image embedding, and on that basis determines whether an image is AI-generated. On Forensics Small, AUROC is on average 39.61% higher than the best training-free method and 5.14% higher than the best training-based method.