Frequency-domain ResNet-34 with stochastic augmentation enhances detection of AI-generated images
Year: 2026
Type: article
Abstract: Nowadays, diffusion-based generators such as Stable Diffusion, Midjourney, or DALL-E have made AI-generated visual content indistinguishable from authentic media, while online platforms often spread such material with inconsistent flagging of synthetic media, contributing to misinformation risks. To address this gap, a lightweight frequency-domain detector combining ResNet-34 with Advanced Stochastic Augmentation (ASA) was proposed, operating exclusively on the Fourier magnitude spectrum. Input images are converted to luminance, transformed via the fast Fourier Transform (FFT), log-scaled, and classified by a ResNet-34 trained on 53,988 class-balanced images from diffusion and GAN sources. ASA combines mobile ISP simulation, variable JPEG compression, blur, scaling, and Gaussian noise injection during training. Evaluation is performed on the SPAI cross-dataset benchmark covering 13 generators, with robustness assessed under compression, resizing, blur, and noise; an ablation study additionally confirms the contribution of ASA. Across 30 independent training runs, ResNet-34-FFT ASA achieves a mean SPAI benchmark AUC of 92.4% (95% CI [92.1%, 92.8%]), compared with 91.0% (SPAI), 85.5% (RINE), and 83.5% (DMID); the best observed run (seed 626) reaches 95.7%. On the internal test set, the best model attains 95.6% accuracy with sustained robustness under common image degradations (AUC $$\ge$$ 91% at JPEG Q = 50).
Source: Scientific Reports
Institution: Tomas Bata University in Zlín
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Funder: Univerzita Tomáše Bati ve Zlíně
Award: G718785734
OpenAlex ID: W7236375180
