The proliferation of highly realistic images generated by advanced AI models has intensified concerns regarding potential misuse and the critical demand for effective AI-generated image detectors. Existing detection techniques face significant hurdles: training-based methods are often computationally intensive and struggle to generalize to novel data domains, while training-free approaches typically fall short in detection accuracy.
To address these limitations, a new research proposes a novel search-based method leveraging data embedding sensitivity within intermediate layers for robust AI-generated image detection. This technique operates by analyzing the similarity between original image embeddings and their perturbed counterparts. Given a collection of real and AI-generated images, the method identifies AI-generated content based on these similarity comparisons.
The efficacy of this proposed method was rigorously evaluated on two extensive benchmarks: GenImage and Forensics Small. The results demonstrate a significant performance improvement across various datasets, surpassing both state-of-the-art training-free and training-based detection techniques. Notably, on the Forensics Small benchmark, the method achieved an average performance gain of 39.61% in AUROC score compared to the best training-free method, and a 5.14% gain over the best training-based method.
