ML Security PapersMIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection
Ruiqi Liu 1,2, Manni Cui 3, Ziheng Qin 1, Zhiyuan Yan 4, Ruoxin Chen 5, Yi Han 6, Zhiheng Li 1, Junkai Chen 1, ZhiJin Chen 1, Kaiqing Lin 7, Jialiang Shen 8, Lubin Weng 1, Jing Dong 1, Yan Wang 9, Shu Wu 1
1 Institute of Automation, Chinese Academy of Sciences
2 School of Advanced Interdisciplinary Sciences, UCAS
Published on arXiv
2602.02222
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
MIRROR achieves 89.6% accuracy across 27 generators on the Human-AIGI benchmark, surpassing both lay users and visual experts, with gains of 2.1% on six standard and 8.1% on seven in-the-wild benchmarks over prior methods.
MIRROR (Manifold Ideal Reference ReconstructOR)
Novel technique introduced
High-fidelity generative models have narrowed the perceptual gap between synthetic and real images, posing serious threats to media security. Most existing AI-generated image (AIGI) detectors rely on artifact-based classification and struggle to generalize to evolving generative traces. In contrast, human judgment relies on stable real-world regularities, with deviations from the human cognitive manifold serving as a more generalizable signal of forgery. Motivated by this insight, we reformulate AIGI detection as a Reference-Comparison problem that verifies consistency with the real-image manifold rather than fitting specific forgery cues. We propose MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. MIRROR projects an input into a manifold-consistent ideal reference via sparse linear combination, and uses the resulting residuals as robust detection signals. To evaluate whether detectors reach the "superhuman crossover" required to replace human experts, we introduce the Human-AIGI benchmark, featuring a psychophysically curated human-imperceptible subset. Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On Human-AIGI, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts, and further approaching the human perceptual limit as pretrained backbones scale. The code is publicly available at: https://github.com/349793927/MIRROR
Key Contributions
- Reformulates AI-generated image detection as a Reference-Comparison problem — projecting inputs to a manifold-consistent ideal reference via a learnable discrete memory bank and using reconstruction residuals as detection signals, avoiding overfitting to generator-specific artifacts.
- MIRROR framework that encodes real-image priors through sparse linear combinations over a discrete memory bank, yielding robust and generalizable detection signals that outperform artifact-driven baselines by 8.1% on in-the-wild benchmarks.
- Human-AIGI benchmark with a psychophysically curated human-imperceptible subset to evaluate 'superhuman crossover', enabling principled comparison against lay users and visual experts across 27 generators.
🛡️ Threat Analysis
Proposes a generalizable AI-generated image detector that verifies content authenticity against a learned real-image manifold — directly addressing output integrity and synthetic content detection, including GAN and diffusion model outputs across 14 benchmarks.