Spatial-Frequency Feature Fusion Network for Lightweight and Arbitrary-Sized JPEG Steganalysis

Current deep learning-based JPEG image steganalysis methods typically rely on decompressed pixels for steganalytic feature extraction, without fully leveraging the inherent information in JPEG images. Additionally, they often face limitations such as large parameter counts and restricted image sizes for detection. In this letter, we propose a spatial-frequency feature fusion network (SF3Net) for lightweight and arbitrary-sized JPEG steganalysis. SF3Net introduces a PReLU activation function and a multi-view convolutional module to capture refined residual features from decompressed pixels, while also integrating original DCT coefficients and quantization tables to extract additional modal features. The spatial-frequency multi-modality features are then fused using a coordinate attention mechanism. And a patch splitting scheme is designed to be compatible with any feature resolution, enabling the detection of arbitrary-sized images with a Swin Transformer block. Experimental results demonstrate that SF3Net outperforms existing methods in detecting both fixed-sized and arbitrary-sized images, while significantly reducing the number of parameters.