Deep Robust Reversible Watermarking.

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Existing Robust Reversible Watermarking (RRW) methods are often complex, computationally expensive, and lack robustness.
• Non-deep learning approaches require task-specific designs for different image distortions.

Purpose of the Study:

• To propose a novel deep learning-based scheme for Robust Reversible Watermarking (DRRW).
• To enhance robustness, reduce computational costs, and improve practical applicability of RRW.

Main Methods:

• Introduced an Integer Invertible Watermark Network (iIWN) for invertible mapping of integer data distributions.
• Employed an encoder-noise layer-decoder framework for adaptive robustness against various distortions via end-to-end training.
• Utilized arithmetic coding for efficient bitstream compression and reversible data hiding, alongside an overflow penalty loss and adaptive weight adjustment.

Main Results:

• Achieved adaptive robustness against diverse distortions without task-specific designs.
• Significantly reduced pixel overflow, improving stego image quality and robustness.
• Demonstrated improved training stability and performance through an adaptive weight adjustment strategy.

Conclusions:

• DRRW effectively addresses key challenges in current RRW methods.
• The proposed scheme significantly advances the practical deployment of robust and reversible watermarking.