Spectral Forensics: Detecting Diffusion-Generated Imagery via $1/f^\beta$ Power Law Violations
IEEE Signal Processing Letters · 2026 · 33 · 1290-1294
DOI: 10.1109/lsp.2026.3672388 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Sridharan Sankaran. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 22 Literaturangaben.
Abstract (Originalsprache)
As generative AI democratizes the creation of hyper-realistic media, the erosion of digital authenticity poses severe risks to societal trust and information security. Current forensic detectors primarily rely on opaque, data-driven deep neural networks, which often function as “black boxes” and struggle to generalize across evolving architectures. In this letter, we propose a robust, physics-based forensic framework designed to restore trustworthiness through transparent spectral analysis. We hypothesize that while diffusion models achieve spatial realism, their iterative denoising process introduces systematic violations of the natural $1/f^\beta$ power-law decay in the frequency domain. By analyzing the azimuthally averaged radial power spectrum, we identify a distinct high-frequency “energy surplus” unique to diffusion synthesis. Rigorous experiments across diverse datasets and architectures (Stable Diffusion v1.5, SDXL, DALL-E 3) confirm this artifact’s universality. A lightweight linear classifier achieves 100% detection accuracy on open-source models and maintains robust zero-shot generalization (82.50%) against proprietary systems. Unlike deep learning baselines, our approach provides verifiable signal-processing evidence, offering a principled solution to distinguish synthetic media from reality. Code used is publicly available for reproduceability.
Zitieren
APA
Sankaran, S. (2026). Spectral Forensics: Detecting Diffusion-Generated Imagery via $1/f^\beta$ Power Law Violations. IEEE Signal Processing Letters, 33, 1290-1294. https://doi.org/10.1109/lsp.2026.3672388
Vancouver
Sankaran S. Spectral Forensics: Detecting Diffusion-Generated Imagery via $1/f^\beta$ Power Law Violations. IEEE Signal Processing Letters. 2026;33:1290-1294. doi:10.1109/lsp.2026.3672388
BibTeX
@article{Sankaran2026,
author = {Sankaran, Sridharan},
title = {Spectral Forensics: Detecting Diffusion-Generated Imagery via $1/f^\beta$ Power Law Violations},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {1290--1294},
doi = {10.1109/lsp.2026.3672388}
}
EUVIMED