AI-generated image detectors are surprisingly easy to mislead... for now

AI-generated image detectors, also known as fake image detectors, have demonstrated remarkable performance across different datasets and generators, achieving superior detection accuracy and generalizability. Considering the structures of existing AI-generated detectors, which perform binary classification, we propose a simple yet effective adversarial attack method, namely Binary Fast Gradient Sign Method (BFGSM), in this paper. We demonstrate that existing AI-generated image detectors are sensitive to subtle and imperceptible distortions, which raises serious safety risk for these models and makes them inadequate for real-world applications. Experimental results show that our proposed attack successfully misleads current AI-generated image detectors, reducing the attack distortion level by 7.72% with negligible impact on the misleading success rate.