Research on Authenticity Identification Methods for Artificial Intelligence-Generated Images
人工智能生成图像真实性识别方法研究

摘要

生成对抗网络、扩散模型和多模态大模型显著降低了高质量图像生成门槛，人工智能生成图像在设计、广告和数 字内容生产中得到广泛应用，同时也增加了虚假新闻、身份冒用、网络诈骗和证据污染风险。针对单一检测特征泛化 能力不足、压缩与重采样后伪影衰减、未知生成模型难以识别等问题，本文提出一种面向开放场景的多证据融合真实 性识别方法。该方法以空间纹理、频域统计、语义一致性和来源凭证为四类核心证据，通过质量感知预处理、分支特 征提取、可信度校准与分层决策完成真实图像、人工智能生成图像和不确定样本的分类，并输出关键区域、异常频段 和来源链等解释信息。研究进一步分析各类检测技术的适用条件，构建由准确性、跨模型泛化性、鲁棒性、校准性和 可解释性组成的评价体系。 

Abstract

Generative adversarial networks, diffusion models, and multimodal large models have significantly lowered the threshold for generating high-quality images. Artificial intelligence-generated images are widely used in design, advertising, and digital content production, while also increasing the risks of fake news, identity impersonation, online fraud, and evidence contamination. To address problems such as insufficient generalization ability of single detection features, artifact attenuation after compression and resampling, and difficulty in identifying unknown generation models, this paper proposes a multi-evidence fusion authenticity identification method for open scenarios. The method takes spatial texture, frequency-domain statistics, semantic consistency, and provenance credentials as four core types of evidence, and completes the classification of real images, artificial intelligence-generated images, and uncertain samples through quality-aware preprocessing, branch feature extraction, credibility calibration, and hierarchical decision-making. It also outputs explanatory information such as key regions, abnormal frequency bands, and provenance chains. The study further analyzes the applicable conditions of various detection technologies and constructs an evaluation system composed of accuracy, cross-model generalization, robustness, calibration, and interpretability.