Scholarlink

We introduce R-SIT, a novel and interpretable architecture for spatial-domain image steganalysis. The framework incorporates a preprocessing stage based on Spatial Rich Model (SRM) filters to amplify steganographic noise, followed by hierarchical Swin Transformer blocks enhanced with Squeeze-and-Excitation (SE) modules. This design achieves a synergy between local detail extraction, global contextual modeling, and channel-wise attention, thereby strengthening the model’s discriminative capacity. We evaluate R-SIT on BOSSBase 1.01 across five widely used steganographic algorithms (WOW, S-UNIWARD, HILL, HUGO, MiPOD). Notably, at 0.4 bpp it achieves 91.05and 87.83visualizations that highlight image regions most influential to classification. These findings underscore the advantages of combining hierarchical Transformers with channel attention and intrinsic explainability mechanisms, offering a robust and transparent solution for steganalysis in the spatial domain.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

R-SIT: A Swin-Transformer-Based Architecture for Spatial Image Steganalysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mariline Catalina Delgado-Martínez,
  • Fabian Alberto Ramirez Rodriguez,
  • Ernesto Guevara-Navarro,
  • Luis Miguel Cardona-Pérez,
  • Santiago Garcia-Herrera,
  • Gustavo Isaza,
  • Reinel Tabares-Soto,
  • Mario Alejandro Bravo-Ortiz

摘要

We introduce R-SIT, a novel and interpretable architecture for spatial-domain image steganalysis. The framework incorporates a preprocessing stage based on Spatial Rich Model (SRM) filters to amplify steganographic noise, followed by hierarchical Swin Transformer blocks enhanced with Squeeze-and-Excitation (SE) modules. This design achieves a synergy between local detail extraction, global contextual modeling, and channel-wise attention, thereby strengthening the model’s discriminative capacity. We evaluate R-SIT on BOSSBase 1.01 across five widely used steganographic algorithms (WOW, S-UNIWARD, HILL, HUGO, MiPOD). Notably, at 0.4 bpp it achieves 91.05and 87.83visualizations that highlight image regions most influential to classification. These findings underscore the advantages of combining hierarchical Transformers with channel attention and intrinsic explainability mechanisms, offering a robust and transparent solution for steganalysis in the spatial domain.