ChaosNet-IE: Neural Network Optimized Chaotic Image Encryption

A high-security image encryption scheme combining an improved Lorenz chaotic system and neural network optimization.

Overview

ChaosNet-IE implements the core ideas from the paper (CN) "基于神经网络优化混沌系统的图像加密算法": generate raw chaotic sequences from an improved Lorenz system, refine them via a BP neural network, then perform pixel permutation and diffusion for robust encryption.

Features

• Neural network assisted enhancement of chaotic sequence quality
• Improved Lorenz dynamics for stronger randomness
• Two-stage processing: permutation + diffusion
• Key derivation via SHA-384(image bytes) -> initial conditions (high key sensitivity)
• Near-ideal entropy (~8.0) and low adjacent-pixel correlation after encryption

Project Structure

ChaosNet-IE/
├── keys/
│   ├── sequences.npz          # Trained/optimized chaotic sequences
│   └── initial_values.txt     # Initial Lorenz parameters
├── output/
│   ├── encrypted.png
│   └── decrypted.png
├── encrypt.py                 # Encryption script
├── decrypt.py                 # Decryption script
├── demo.py                    # One-click demo
├── lena.png                   # Test image
└── README.md / README_CN.md

Quick Start

Install Dependencies

pip install numpy opencv-python

One-Click Demo

python demo.py

Performs: encryption → decryption → statistics output.

Encrypt Only

python encrypt.py

Outputs: output/encrypted.png, keys/sequences.npz, keys/initial_values.txt

Decrypt Only

python decrypt.py

Outputs: output/decrypted.png

Algorithm Outline

1. Improved Lorenz system generates chaotic sequences (with warm-up iterations)
2. Sequence normalization to [0,1]
3. BP neural network (≈10 hidden neurons) refines sequence quality
4. Pixel permutation (index scrambling)
5. Block-wise diffusion (XOR / additive mapping)
6. Produce final cipher image

Improved Lorenz Variant

dx/dt = a(y - x)
dy/dt = bx - xz + y
dz/dt = 200x^2 + 0.01·e^(xy) - cz

Conceptual Flow

graph LR
    A[Input Image] --> B[Key Derivation]
    B --> C[Improved Lorenz]
    C --> D[Neural Network]
    D --> E[Permutation]
    E --> F[Diffusion]
    F --> G[Cipher Image]

Loading

Security Indicators (Example)

Metric	Plain	Encrypted	Note
Entropy	~7.0	~8.0	Near ideal randomness
Adjacent Corr.	High	<0.001	Spatial decorrelation
Pixel Change Rate	0%	>99%	Strong diffusion
Histogram	Structured	Uniform	Resists frequency analysis

Visualization

Plain	Encrypted	Decrypted

Key & Sequence Processing

• SHA-384 over raw image bytes -> parse into initial conditions
• 1000 warm-up iterations ensure chaotic regime
• BP network (tanh hidden + linear output) fits/refines series
• Derived indices drive permutation & diffusion vectors

Usage Notes

1. Keep keys/ files safe; they are required for decryption
2. Current implementation targets grayscale images (RGB extension: per-channel processing)
3. Changing the input image requires regenerating key materials

Reference (Chinese Paper)

Paper: 基于神经网络优化混沌系统的图像加密算法 (计算机系统应用) Link: https://www.c-s-a.org.cn/1003-3254/7578.html

Roadmap (Optional Ideas)

• Color image direct support
• GPU acceleration for sequence generation
• Configurable network depth
• Benchmark suite (speed / entropy / NPCR / UACI)

License

Academic / research only. Not intended for commercial deployment.

---

For the Chinese version, click the 中文 badge at the top.