Spectral Tail Auxiliary Learning for AI-Generated Image Detection

2026-05-21

AI Summary

As the perceptual gap between AI-generated and real images continues to narrow, existing detection methods face a generalization bottleneck due to their limited understanding of stable discriminative cues in the frequency domain. This work identifies and formally names the “spectral tail elevation” phenomenon—characterized by anomalous energy amplification in the ultra-high-frequency region of the one-dimensional radial log-power spectrum of synthetic images—and attributes it to nonlinear harmonic accumulation inherent in generative models. Leveraging this insight, the authors propose a zero-inference-overhead frequency-domain auxiliary learning framework that transfers spectral tail cues to a spatial-domain detector via teacher–student knowledge distillation. The method demonstrates strong cross-generator, cross-distribution, and real-world generalization across nine public benchmarks, requiring no additional computation during inference.

Abstract

As generative image models evolve rapidly, the perceptual gap between generated and real images continues to narrow, making AI-generated image detection increasingly challenging. Many existing methods exploit frequency-domain cues for detection, typically described as frequency-domain artifacts or high-frequency discrepancies. However, the specific and recurring spectral regularities remain insufficiently understood and characterized. In this paper, we systematically analyze the one-dimensional radial log-power spectra of real and generated images. We find that generated images do not necessarily exhibit higher or lower energy across the entire spectrum or high-band range. Instead, their spectra deviate from the power-law decay and show an anomalous uplift in the ultra-high-frequency tail. We term this phenomenon spectral tail uplift. We further attribute this phenomenon to nonlinear harmonic accumulation in trained generative models, suggesting that it can serve as a structural cue across generative architectures. Based on this observation, we propose Spectral Tail Auxiliary Learning (STAL), a frequency-domain auxiliary supervision framework for generalizable AI-generated image detection. STAL transfers spectral-tail cues from a tail-aware frequency teacher to a spatial detector during training, while all frequency-domain modules are discarded at inference time. Consequently, STAL introduces no inference overhead. Extensive experiments on 9 public datasets show that STAL achieves strong generalization and stability across generators, data distributions, and real-world scenarios.

Problem

• AI-generated image detection
• spectral regularities
• frequency-domain artifacts
• power-law decay
• ultra-high-frequency tail

Innovation

• spectral tail uplift
• frequency-domain artifacts
• auxiliary learning
• AI-generated image detection
• power-law decay