Steganalysis Network for Weak Steganographic Signal Extraction and Enhancement.

Area of Science:

• Computer Science
• Information Security
• Digital Forensics

Background:

• Digital image steganalysis aims to detect hidden data within cover images.
• Spatial domain steganography introduces subtle modifications (±1 pixel value) that are challenging for standard methods.
• Existing convolutional neural networks often overlook the specific challenges of detecting these weak signals.

Purpose of the Study:

• To develop an advanced steganalysis network capable of effectively extracting and enhancing weak steganographic signals.
• To improve the accuracy and generalization of spatial domain steganalysis algorithms.

Main Methods:

• Proposed a novel preprocessing structure: learnable filter constrained by high-pass prior (LFCHP).
• Introduced a second-order signal auxiliary branch (SSAB) to mitigate signal suppression during convolution.
• Developed a new pooling method, SoftPool, to minimize signal loss during downsampling.
• Integrated these components into a new steganalysis network named WSERNet.

Main Results:

• WSERNet achieved accuracy improvements of 1.08-2.96% compared to state-of-the-art spatial-domain steganalysis algorithms.
• The proposed method demonstrated superior performance across three steganographic schemes and four embedding rates.
• Experiments confirmed excellent generalization capabilities across different steganography techniques.

Conclusions:

• The novel components (LFCHP, SSAB, SoftPool) significantly enhance WSERNet's ability to detect weak steganographic signals.
• WSERNet represents a significant advancement in spatial domain steganalysis, offering improved accuracy and robustness.