Lossless Image Steganography Based on Invertible Neural Networks

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Traditional image steganography methods often prioritize visual similarity over secret information recovery accuracy.
• Existing techniques may compromise the integrity of hidden data during the steganography process.

Purpose of the Study:

• To propose an image steganography method utilizing invertible neural networks for enhanced invisibility, security, and lossless data recovery.
• To introduce a mapping module for compressing embedded information, improving stego-image quality and anti-detection capabilities.

Main Methods:

• Developed a steganography scheme based on invertible neural networks (INNs).
• Integrated a mapping module to compress secret information before embedding.
• Converted secret information into a binary sequence for embedding via INN forward operations.
• Recovered secret information using the inverse operations of the INNs.

Main Results:

• Achieved high invisibility and security in stego images.
• Demonstrated lossless recovery of secret information.
• Improved stego-image quality and anti-detection performance through information compression.
• Experimental results showed competitive performance in visual quality and security.

Conclusions:

• The proposed invertible neural network-based image steganography method offers superior performance in terms of invisibility, security, and data recovery.
• The integration of a mapping module further enhances stego-image quality and resistance to detection.
• The method guarantees 100% accuracy in secret information extraction, addressing a key limitation of prior approaches.