AOT-PixelNet: Lightweight and interpretable detection of forged images via adaptive orthogonal transform

Abstract

Generative image detection faces persistent challenges in terms of generalization and interpretability, limiting its reliability in complex scenarios. To address these issues, we propose AOT-PixelNet, a lightweight and interpretable detection framework that integrates an Adaptive Orthogonal Transform (AOT) module with a streamlined 1 × 1 convolution-based PixelNet architecture. The AOT module leverages diverse orthogonal transforms, such as FFT and DCT, to extract informative frequency-domain features, thereby enhancing sensitivity to medium- and high-frequency artifacts. Meanwhile, PixelNet minimizes parameter count (only 0.98 million) while effectively capturing cross-channel inconsistencies and mitigating overfitting. Experimental evaluations on multiple unseen GAN and diffusion-based datasets demonstrate that AOT-PixelNet achieves superior performance with minimal computational cost. Specifically, it outperforms the NPR method by 0.6% and 11.76% on the ForenSynths and GenImage datasets, respectively, validating the framework’s robustness, effectiveness, and interpretability.