Sensitivity of deep learning applied to spatial image steganalysis.

Area of Science:

• Computer Science
• Cybersecurity
• Digital Forensics

Background:

• Traditional spatial image steganalysis is increasingly adopting deep learning (DL) methods.
• Deep learning models, particularly convolutional neural networks (CNNs), integrate feature extraction and classification for improved detection accuracy.
• Existing research primarily focuses on novel network architectures to boost steganalysis performance.

Purpose of the Study:

• To evaluate the impact of preprocessing and database partitioning on CNN-based steganalysis performance.
• To investigate the influence of various normalization ranges, database splits, and activation functions on novel steganalysis networks.
• To analyze activation maps and provide recommendations for robust experimental design in DL steganalysis.

Main Methods:

• Testing novel steganalysis networks (Xu-Net, Ye-Net, Yedroudj-Net, SR-Net, Zhu-Net, GBRAS-Net).
• Employing diverse combinations of image/filter normalization, database splits, and preprocessing activation functions.
• Analyzing activation maps to understand network behavior.

Main Results:

• Steganalysis system performance is highly sensitive to variations in preprocessing and data handling stages.
• Different network architectures and experimental configurations yield varying detection accuracies.
• Thorough documentation and reporting of experimental parameters are crucial for reproducibility and comparison.

Conclusions:

• The effectiveness of deep learning-based steganalysis is significantly influenced by methodological choices in data preprocessing and partitioning.
• Researchers must meticulously report experimental details to ensure the validity and comparability of their findings.
• Standardized experimental protocols are recommended for advancing the field of DL steganalysis.