Robust data hiding for JPEG images with invertible neural network

Area of Science:

• Digital Image Processing
• Information Security
• Machine Learning

Background:

• JPEG compression introduces significant distortions, challenging secure data extraction from stego images.
• Existing data hiding methods struggle with robustness against lossy compression.

Purpose of the Study:

• To develop an end-to-end robust data hiding scheme specifically for JPEG images.
• To enhance the security and resilience of secret message extraction from compressed images.

Main Methods:

• Utilizing an invertible neural network (INN) for bi-directional embedding and extraction of secret messages.
• Operating on quantized Discrete Cosine Transform (DCT) coefficients for inherent compression robustness.
• Incorporating a JPEG compression attack module to train the network for recovery from compressed images.

Main Results:

• The proposed scheme demonstrates strong robustness against lossy JPEG compression.
• Significant improvements in security compared to existing methods were observed.
• The method maintains high image quality and capacity while enhancing data hiding security.

Conclusions:

• The INN-based approach offers intrinsic robustness against JPEG compression artifacts.
• This novel scheme effectively addresses the challenges of data hiding in compressed images.
• The method provides a secure and high-capacity solution for steganography in JPEG images.