A new JPEG image steganalysis technique combining rich model features and convolutional neural networks

Area of Science:

• Computer Science
• Digital Image Processing
• Cybersecurity

Background:

• Traditional steganalysis methods for adaptive steganography often rely on rich models and ensemble classifiers.
• These methods face challenges with high-dimensional data in steganalysis.
• Adaptive steganography poses unique detection challenges due to its dynamic nature.

Purpose of the Study:

• To propose a new steganalysis method for JPEG images using convolutional neural networks (CNNs).
• To address the high-dimensionality problem in steganalysis.
• To improve the detection performance for adaptive steganography.

Main Methods:

• A novel CNN-based steganalysis method is developed for JPEG images.
• The method enhances existing rich models by incorporating discrete cosine transform (DCT) basis functions of varying sizes.
• Extracted features are processed through a neural network's convolutional and fully connected layers for classification.

Main Results:

• The proposed CNN method demonstrates effective steganalysis for JPEG images.
• It requires fewer training samples compared to traditional methods.
• Experimental results show superior classification performance over original ensemble classifiers.

Conclusions:

• Convolutional neural networks offer a viable and efficient solution for JPEG image steganalysis.
• The integration of DCT basis functions enhances feature extraction capabilities.
• The CNN-based approach provides improved detection accuracy and efficiency for adaptive steganography.