SCADET: A detection framework for AI-generated artwork integrating dynamic frequency attention and contrastive spectral analysis

Abstract

With the rapid development of generative AI technology, AI-generated images pose significant challenges for authenticity verification and originality validation. This paper proposes SCADET, a novel detection framework that integrates Dynamic Frequency Attention Network (DFAN) and Contrastive Spectral Analysis Network (CSAN). DFAN adaptively analyzes image frequency domain features and dynamically adjusts attention for different artistic styles, while CSAN establishes discriminative feature spaces through contrastive learning to enhance cross-model generalization capabilities. Comprehensive experiments on the AI-ArtBench dataset demonstrate that SCADET achieves AUC values of 0.962 and 0.801 in full image and local image detection tasks respectively, representing substantial improvements of 30.5% and 34.4% over baseline methods. Cross-model evaluation shows that the framework maintains stable performance across various generation techniques, with an average accuracy of 0.81 and low variance. Ablation studies validate the effectiveness of both DFAN and CSAN components. These results advance the field of AI-generated content detection and provide valuable insights for addressing authenticity challenges in digital media applications.