Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection

2025-06-01

AI Summary

AIGC image detectors exhibit poor generalization across unseen generative models, primarily because they rely on non-robust shortcut features—such as initial noise priors—rather than authentic generation artifacts. This work identifies and mitigates this latent-variable prior bias via **Orthogonal Manifold Adversarial Training (OMAT)**: under fixed conditional constraints, it optimizes the initial latent variables of diffusion models to generate reproducible, manifold-preserving adversarial examples. We further introduce **GenImage++**, a novel multi-source benchmark for cross-generator evaluation. Our approach integrates latent-variable optimization, adversarial training, and cross-generator assessment within ResNet50- and CLIP-based detector architectures. Experiments demonstrate substantial improvements in cross-generator robustness, achieving state-of-the-art generalization performance on outputs from cutting-edge models including Flux.1 and SD3—without modifying the detector’s architecture.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

Current AIGC detectors often achieve near-perfect accuracy on images produced by the same generator used for training but struggle to generalize to outputs from unseen generators. We trace this failure in part to latent prior bias: detectors learn shortcuts tied to patterns stemming from the initial noise vector rather than learning robust generative artifacts. To address this, we propose On-Manifold Adversarial Training (OMAT): by optimizing the initial latent noise of diffusion models under fixed conditioning, we generate on-manifold adversarial examples that remain on the generator's output manifold-unlike pixel-space attacks, which introduce off-manifold perturbations that the generator itself cannot reproduce and that can obscure the true discriminative artifacts. To test against state-of-the-art generative models, we introduce GenImage++, a test-only benchmark of outputs from advanced generators (Flux.1, SD3) with extended prompts and diverse styles. We apply our adversarial-training paradigm to ResNet50 and CLIP baselines and evaluate across existing AIGC forensic benchmarks and recent challenge datasets. Extensive experiments show that adversarially trained detectors significantly improve cross-generator performance without any network redesign. Our findings on latent-prior bias offer valuable insights for future dataset construction and detector evaluation, guiding the development of more robust and generalizable AIGC forensic methodologies.

Problem

• Detectors fail to generalize across unseen AIGC generators
• Latent prior bias causes reliance on non-robust noise patterns
• Need for adversarial training to improve cross-generator detection

Innovation

• On-Manifold Adversarial Training (OMAT) optimizes latent noise
• Generates on-manifold adversarial examples for robust detection
• Improves cross-generator performance without network redesign