Evaluating Frequency and Noise-Residual Features for Robust AI-Generated Image Detection

Abstract Adding hand-crafted forensic features, such as Fourier spectra or high-pass noise residuals, is a common way to help detectors distinguish AI-generated images from photographs. Yet it remains unclear whether these extra inputs actually help when a detector encounters generators it has not seen during training. We examine this question through a controlled ablation using MobileNetV2 and four input configurations: RGB alone, RGB with an FFT channel, RGB with an SRM residual channel, and RGB with both features. Experiments use 19,964 images from CommunityForensics-Small, representing 269 generators. We hold out 54 generators, repeat training with three random seeds, and test the models under JPEG compression and down-scaling. RGB alone gives the best clean-image accuracy: 94.77% on known generators and 93.91% on unseen ones. The three feature-augmented configurations reach 92.67%–93.56% and 93.15%–93.24%, respectively. None outperforms RGB significantly in any of the 18 comparisons after Holm correction of exact McNemar tests. The residual channel is especially vulnerable to resizing: at 50% down-scaling, accuracy on unseen generators falls from 87.31% for RGB to 80.54% with the residual alone and 82.33% with both additional channels. Importantly, 53 of the 54 held-out generators belong to the latent-diffusion family, while the single held-out GAN is detected in only 31%–37% of cases. The findings therefore support generalization mainly within a generator family, not across architectures. In this setting, the extra channels increase complexity without a measurable improvement.