Hang Yang

Hang Yang

he/him

PhD candidate in Mathematics at UMass Amherst. Researching machine learning, robotics, and deep learning — including mobile motion capture with robot dogs, invertible neural networks for image steganography, and vision transformers for medical AI.

PRIS: Practical robust invertible network for image steganography

Published in Engineering Applications of Artificial Intelligence, 2024

We present PRIS, a practical robust invertible network for image steganography. Existing methods suffer from poor robustness when container images undergo distortions (e.g., Gaussian noise, JPEG compression) and fail to handle rounding errors arising from the mismatch between deep learning precision (32-bit) and standard image bit-depth (8-bit).

PRIS addresses these issues with:

  • A pre-enhance module that restores distorted container images before extraction
  • A post-enhance module that refines the extracted secret image
  • A gradient approximation function (GAF) to handle the non-differentiable rounding operation
  • A 3-step training strategy to preserve invertibility while learning enhancement

Experimental results show PRIS achieves an average PSNR of 34.28/32.96 dB on container/secret pairs under 5 different attack types, outperforming state-of-the-art robust steganography methods.

Code available at https://github.com/yanghangAI/PRIS.

Recommended citation: Yang, H., Xu, Y., Liu, X., & Ma, X. (2024). "PRIS: Practical robust invertible network for image steganography." Engineering Applications of Artificial Intelligence, 133, 108419.
Download Paper