Imagine hiding a secret message inside an image, transmitting it, and then extracting that exact message without losing a single bit. Now, imagine being able to restore the original cover image to its exact mathematical state, pixel for pixel. 🧩

In this article, I will walk through my implementation of a research-backed architecture that uses Bayesian Neural Networks and Modulo-256 arithmetic to achieve 100% reversible steganography.

---

📜 Academic Attribution

First and foremost, this work is an implementation and exploration of the research paper:

"Bayesian Neural Networks for Reversible Steganography"
Author: Ching-Chun Chang
Journal: IEEE Access, 2022
DOI: 10.1109/ACCESS.2022.3159911

All architectural foundations for the uncertainty-driven embedding, including the Bayesian Neural Network predictor and the variance-sorted difference expansion, are attributed to the original author.

---

📌 The Challenge: The Reversibility Problem

Most traditional steganography permanently alters the cover image. Reversible steganography aims to fix this, but it faces a massive enemy: pixel overflow and underflow (clipping).

If a pixel is purely white (255) and you add 1 to hide a secret bit, the image format clips it back to 255. When the receiver tries to read it, the math is broken, the secret bit is corrupted, and the image cannot be restored. 💥

---

🧬 Why Bayesian Neural Networks?

The core innovation here is uncertainty estimation. Instead of just guessing pixel values, the network outputs two things: the predicted pixel intensity and its own uncertainty (variance).

By predicting which pixels it is highly confident about, the network creates a variance map. We can then hide our secret message exclusively in the areas where the network is most certain. This ensures the visual distortion is kept to an absolute minimum. 🧠

---

🔄 The Modulo-256 Magic

To completely defeat the clipping problem, this implementation uses Modulo-256 arithmetic. If a pixel at 255 gets a +1 modification, it rolls over to 0 instead of clipping.

Because the receiver uses a checkerboard pattern to share the exact same neural network context, they can calculate the modulo difference and perfectly restore both the secret bit and the original pixel.

---

🧮 The Mathematical Framework

The system optimizes a custom uncertainty-weighted loss function. Instead of standard mean squared error, the loss penalizes errors based on the model's confidence:

$$Loss = \frac{1}{N} \sum \left( \frac{(y - \hat{y})^2}{\sigma^2} + \lambda \log(\sigma^2) \right)$$

For the embedding process, the residual \epsilon (the difference between the true pixel and the prediction) is modulated with the secret bit m:

$$\epsilon^{\prime} = (2 \times \epsilon) + m \pmod{256}$$

---

📊 Experimental Results (The Trial Run)

I trained the Bayesian predictor for 200 epochs on a dataset of standard test images. The payload size was set to approximately 0.3 bits per pixel (bpp).

1. Visual Performance & Reversibility

The system achieved 100% extraction success. Every single bit was recovered, and the images were restored flawlessly. For smooth images like Lena and Sailboat, the PSNR reached above 50 dB, meaning the stego image is visually indistinguishable from the original.

Below are the visual results across different textures, demonstrating the system's ability to maintain imperceptibility (ranked from highest to lowest PSNR):

1. Sailboat

• MSE: 0.5237
• PSNR: 50.94 dB

2. Lena

• MSE: 0.6204
• PSNR: 50.20 dB

3. Airplane

• MSE: 0.7910
• PSNR: 49.15 dB

4. Baboon

• MSE: 1.7721
• PSNR: 45.65 dB

5. Splash

• MSE: 1.9156
• PSNR: 45.31 dB

6. Peppers

• MSE: 7.5769
• PSNR: 39.34 dB

2. Training Convergence

The training curve shows the stabilization of the Bayesian loss across the 200 epochs.

The loss function descends smoothly without exploding gradients, indicating that the model successfully learns to predict pixel values while accurately estimating its uncertainty bounds. 📉

---

🏁 Conclusion

Reversible steganography is a fascinating intersection of deep learning and cryptography. By combining Bayesian uncertainty with clever modular arithmetic, we can achieve high-capacity hiding with zero data loss. 🛡️

Check out the repository here: Anjasfedo/bayesian_neural_network_reversible_steganography

The repository is structured cleanly in PyTorch, allowing you to train the network on your own datasets and test the limits of reversible embedding. 💻

Have you ever worked with Bayesian Networks for tasks outside of standard classification? Let me know in the comments! 👇