Deep steganographic approach for reliable data hiding using convolutional neural networks and adaptive loss...

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Steganography is crucial for secure data transmission.
• Existing methods face challenges in payload capacity and robustness.
• Deep learning offers potential for advanced steganographic solutions.

Purpose of the Study:

• To propose a novel deep steganography framework for robust data hiding.
• To optimize the framework using a specialized loss function.
• To evaluate the framework's performance against state-of-the-art methods.

Main Methods:

• A three-layered Convolutional Neural Network (CNN) architecture was developed, including preparation, hiding, and revealing networks.
• The preparation network utilized filters for edge feature extraction.
• The hiding network employed adaptive embedding, optimized with four loss functions, particularly Loss Function 3 (LF 3).

Main Results:

• Loss Function 3 (LF 3) achieved a payload of 3-5 bits per pixel and improved Peak Signal-to-Noise Ratio (PSNR).
• The framework demonstrated robustness against Gaussian noise, cropping, and rotation.
• Low detection rates were observed across various steganalysis techniques (histogram, statistical, CNN-based).

Conclusions:

• The proposed deep steganography framework with LF 3 offers a superior balance between payload and robustness.
• It outperforms Generative Adversarial Network (GAN)-based methods in terms of efficiency and performance.
• This advancement holds promise for secure data hiding applications, including medical imaging.