基于通道–空间自适应与多残差环注意力机制的人工智能生成图像检测

AI-Generated Image Detection Based on Channel-Spatial Adaptation and Multi-Residual Ring Attention Mechanism

摘要: 针对生成式人工智能快速发展导致的AI生成图像真实性验证难题，以及现有检测方法依赖单模态特征、泛化能力有限的问题，旨在提升对未知生成模型和复杂后处理操作的检测准确率与鲁棒性。首先设计了通道–空间注意力自适应残差模块，通过双重注意力机制和自适应融合策略增强图像的高频噪声、边缘梯度和频域变换关键特征；再引入残差环增强的多头注意力机制，在标准多头注意力计算之后串联残差环，对注意力输出进行迭代式非线性后处理，提高特征交互的表示复杂度；最后进行多层次特征的稳定融合。在GenImage数据集上的实验结果表明，所提方法对七种主流生成模型生成的图像检测准确率达到99.67%，显著优于现有最先进方法。该方法展现了优越的泛化能力和鲁棒性，能够有效应对未知生成模型与复杂后处理操作，为AI生成图像检测提供了高效可靠的技术途径。

Abstract: To address the challenge of authenticating AI-generated images arising from the rapid development of generative artificial intelligence, and to overcome the limitations of existing detection methods that rely on single-modal features and exhibit limited generalization capability, this study aims to improve detection accuracy and robustness against unknown generative models and complex post-processing operations. We first design a channel-spatial attention adaptive residual module, which enhances key features-including high-frequency noise, edge gradients, and frequency-domain transform characteristics-through a dual-attention mechanism and an adaptive fusion strategy; then we introduce a residual ring-enhanced multi-head attention mechanism that appends a residual ring after the standard multi-head attention computation to perform iterative nonlinear post-processing on the attention output, thereby increasing the representational complexity of feature interactions; and finally we perform stable fusion of multi-level features. Experimental results on the GenImage dataset demonstrate that the proposed method achieves a detection accuracy of 99.67% on images generated by seven mainstream generative models, significantly outperforming existing state-of-the-art methods. The proposed method exhibits superior generalization capability and robustness, effectively handling unknown generative models and complex post-processing operations, thus providing an efficient and reliable technical approach for AI-generated image detection.