Intermediate Representations are Strong Training-Free AI-Generated Image Detectors

ICLR 2026 Conference Submission5881 Authors

Keywords: Fake Image Detection, Adversarial Training, Computer Vision

Abstract: The rapid advancement in generative AI models has enabled the creation of photorealistic images. At the same time, there are growing concerns about the potential misuse and dangers of generated content, as well as a pressing need for effective AI-generated image detectors. However, current training-based detection techniques are typically computationally costly and can hardly be generalized to unseen data domains, while training-free methods fall short in detection performance. To bridge this gap, we propose a training-free method employing data embedding sensitivity in intermediate layers to detect AI-generated images. Given a set of real and AI-generated images, our method scans through the detection performance in the composite configuration space of intermediate layer, perturbation type, and severity level to identify the best configuration for detection. We examine the proposed method on two comprehensive benchmarks: GenImage and DF40. Our method exhibits improved performance across different datasets compared to both training-free and training-based state-of-the-art methods. On average, our method outperforms the best training-free/training-based methods on the GenImage benchmark by 16.1%/4.9% and on the DF40 benchmark by 14.5%/8.7% in AUROC score. We release the code at https://anonymous.4open.science/r/Intermediate-Public-D256.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 5881