.png){kind=icon}
.png){kind=icon}
Return
Synthetic Image Detection Based on Multi-Level Feature Fusion and Data Augmentation
.png){kind=icon}
OA
AI
G
Y
J
DOI：10.1016/j.icte.2026.05.018.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the rapid progress of generative artificial intelligence (AI), AI-generated images have become increasingly realistic and diverse, posing serious challenges to information security and media credibility. However, conventional synthetic image detection (SID) methods often rely on generator-specific artifacts, leading to performance degradation on unseen generators. Recent CLIP-based approaches, such as UFD and RINE, have shown strong potential by using large-scale pretrained visual representations. Nevertheless, UFD mainly relies on final-layer CLIP features and may miss useful intermediate-layer information, whereas RINE adopts a more complex intermediate-feature fusion design with more trainable parameters. To address this issue, we propose SNIFF (SyNthetic Image Detection Based on Multi-Level Feature Fusion), a simple and parameter-efficient CLIP-based SID framework that integrates SAFE (Simple Preserved and Augmented Features) preprocessing with multi-level intermediate feature fusion. SNIFF uses a frozen CLIP Vision Transformer (CLIP-ViT) backbone to extract class-token features from multiple intermediate layers, fuses them through element-wise averaging, and applies a streamlined two-layer linear classifier for authenticity prediction. Experiments on the large-scale GenImage benchmark show that SNIFF achieves 86.6% average precision (AP) and 79.5% average accuracy (ACC), outperforming UFD and RINE. Ablation and efficiency analyses further show that SNIFF improves detection performance while using only 0.66M trainable parameters, providing a practical lightweight solution for CLIP-based SID.
Keywords:
Synthetic image detection
CLIP-ViT
Multi-level feature fusion
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4.2
Papers:
960
Citations:
2.5K
Organization
No organization information available
Cited Papers