[Paper Review] Multi-Image Steganography Using Deep Neural Networks

Abhishek Das, Japsimar Singh Wahi+2 more|arXiv (Cornell University)|Jan 2, 2021

Advanced Steganography and Watermarking Techniques3 references29 citations

TL;DR

The paper develops a deep neural network framework to embed multiple secret images into a single same-size cover image and recover them with minimal loss, extending single-image steganography to multi-image scenarios.

ABSTRACT

Steganography is the science of hiding a secret message within an ordinary public message. Over the years, steganography has been used to encode a lower resolution image into a higher resolution image by simple methods like LSB manipulation. We aim to utilize deep neural networks for the encoding and decoding of multiple secret images inside a single cover image of the same resolution.

Motivation & Objective

• Motivate and formalize the problem of hiding multiple full-size secret images inside a single cover image of the same resolution.
• Leverage and extend prior deep steganography architectures to enable multi-image embedding without perceptible alterations to the cover.
• Evaluate the approach on Tiny ImageNet and demonstrate retrieval of multiple secrets with acceptable reconstruction loss.
• Explore network architectures with multiple prep and reveal networks inspired by prior work to support multi-image decoding.

Proposed method

• Adopt an encoder–decoder framework combining prep networks for each secret image with a central hiding network to produce an encoded cover image.
• Concatenate prep outputs for all secret images with the cover and process through a hiding network to generate the container image.
• Use multiple reveal networks (one per secret image) to separately decode each hidden image from the container image.
• Add Gaussian noise to encoder output to distribute secret information beyond LSB manipulation and improve hiding robustness.
• Train with a loss that combines cover fidelity and per-secret reconstruction losses, using a weighted sum similar to L = λc||C−C′||^2 + Σλs||Si−S′i||^2.
• Employ 750–750+ epochs on Tiny ImageNet (64x64) with Adam optimizer and a staged learning-rate schedule; freeze reveal network weights during end-to-end training to stabilize decoding.

Experimental results

Research questions

• RQ1 Can multiple full-size secret images be reliably embedded and later recovered from a single cover image of the same resolution without obvious perceptual changes?
• RQ2 How does the number of hidden secret images affect reconstruction quality and encoded-cover fidelity under a unified loss framework?
• RQ3 What architectural configurations (multiple prep/reveal networks) enable scalable multi-image steganography while maintaining decoding accuracy?
• RQ4 What are practical training settings (loss weights, learning rate schedules, noise injection) that balance cover visual integrity and secret-image quality?

Key findings

• The approach demonstrates embedding and decoding up to three secret images in one 64×64×3 cover image with the encoded cover remaining visually similar to the original.
• After 750 training epochs, reported losses include: entire setup 182053.70, secret1 51495.24, secret2 39911.16, secret3 39337.07, and cover 51310.23.
• The encoded cover image preserves visual fidelity while decoded secret images show minimal loss relative to originals, though encoded covers become more lossy as the number of hidden images increases.
• The study extends Baluja (2017) single-image steganography with multiple reveal networks (one per secret image) and demonstrates feasibility of multi-image embedding without downsampling of secret images.
• Three-secret-image experiments indicate successful retrieval of all secrets, with increased but manageable losses as more secrets are added.