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

Modern steganalysis achieves strong performance on controlled benchmarks, yet existing datasets fail to capture the diversity of real-world JPEG images. In particular, widely used benchmarks employ fixed resolutions, limited quantization tables, and a single chroma subsampling configuration, whereas real JPEG data exhibit heterogeneous resolutions, quantization tables, and subsampling schemes. To bridge this gap, we propose a Real-World Evaluation Protocol built from large-scale web images, which are JPEG compressed using quantization tables derived from JPEG-header statistics and realistic chroma subsampling settings. Our analysis reveals substantial performance degradation under both quantization-table and subsampling mismatches, indicating that practical distribution shifts arise jointly from these factors. To address this, we introduce quantization-table targeting and extend it to subsampling through a subsampling-targeting evaluation protocol, demonstrating that explicit alignment reduces JPEG-parameter mismatch within the proposed protocol. We further instantiate this protocol with RWNet, a JPEG-domain detector that processes QTable-aware dequantized DCT inputs using JPEG-structured preprocessing and SPP-based feature aggregation; separately, the framework uses block-aligned ROI extraction to select local evidence from arbitrary-size JPEG images without global resizing. Together, these components provide a framework for evaluating JPEG-side distribution shifts and studying metadata-guided alignment under the compression conditions explicitly covered by RWSteg.