JPEG-aware steganalysis: explicit frequency modeling with dequantized DCT for real-world JPEG images.

Area of Science:

• Digital Image Forensics
• Computer Vision
• Information Security

Background:

• Current steganalysis benchmarks lack real-world JPEG image diversity.
• Existing datasets use fixed resolutions, quantization tables, and chroma subsampling, unlike real-world data.

Purpose of the Study:

• To propose a Real-World Evaluation Protocol for JPEG steganalysis.
• To address performance degradation caused by JPEG parameter mismatches.
• To introduce metadata-guided alignment for improved steganalysis.

Main Methods:

• Developed a Real-World Evaluation Protocol using large-scale web images with realistic JPEG compression parameters.
• Introduced quantization-table targeting and subsampling-targeting evaluation.
• Instantiated the protocol with RWNet, a JPEG-domain detector with JPEG-structured preprocessing and block-aligned ROI extraction.

Main Results:

• Significant performance degradation observed when JPEG quantization tables and chroma subsampling differ from training data.
• Explicit alignment of quantization tables and subsampling reduces JPEG-parameter mismatch.
• RWNet demonstrates effective evaluation of JPEG-side distribution shifts and metadata-guided alignment.

Conclusions:

• Real-world JPEG image diversity, particularly quantization tables and subsampling, significantly impacts steganalysis performance.
• The proposed evaluation protocol and RWNet provide a framework for robust JPEG steganalysis.
• Metadata-guided alignment is crucial for adapting steganalysis models to diverse JPEG compression conditions.