Spectral-artifact image statistic (azimuthal FFT power ridge), opt-in

[mgiovani]

Why

Upsampling in GANs and diffusion decoders leaves periodic artifacts in the frequency domain. Three papers show a hand-crafted spectral statistic separates real from synthetic without a trained network; the threshold is the only tuned part.

Sources

• Durall et al. 2019, Unmasking DeepFakes with simple features (azimuthal average of the FFT power spectrum, basic classifier): https://arxiv.org/abs/1911.00686
• Frank et al. 2020, Leveraging frequency analysis for deep fake image recognition (DCT grid artifacts from upsampling, consistent across GAN architectures): https://arxiv.org/abs/2003.08685
• Corvi et al. 2023, Intriguing properties of synthetic images (diffusion and VQ-GAN output also show Fourier peaks and autocorrelation patterns; mid-high radial and angular spectral power differs from real): https://arxiv.org/abs/2304.06408
• Survey: https://arxiv.org/abs/2502.15176. Benchmark saturation warning: https://arxiv.org/abs/2406.19435

Scope

• One rule, Tier B, off by default: decode to grayscale, FFT, azimuthal average of the power spectrum, flag when the high-frequency energy ridge exceeds a threshold.
• Threshold tuned on a labelled real/synthetic corpus and cited in the doc comment. Known false positives: upscaled or resized real photos. Ships on by default only after that rate is measured.
• Dependencies: png and jpeg-decoder (or image with default features off) and rustfft. Each needs the PR-body line AGENTS.md requires saying what std could not do.
• Depends on the image metadata issue for Lang::Image and ImageContext.