zameer-Gimba/steganalysis-hybridized-lu-cnn
Record type:
project
Creator:
zam
Host:
This project explores a dynamic hybridization of LU-based activation functions (ELU, ReLU, and LeakyReLU) within a Convolutional Neural Network (CNN) architecture for Steganalysis. The work was developed during my SIWES (Industrial Training) at NITDA ICT Hub, Federal University Kashere (FUK), Gombe State.
# Hybrid LU Activation Function in CNN for Steganalysis
## Overview
This project explores a **dynamic hybridization of LU-based activation functions**
(ELU, ReLU, and LeakyReLU) within a Convolutional Neural Network (CNN) architecture
for **steganalysis**.
The work was developed during my **SIWES (Industrial Training)** at
**NITDA ICT Hub, Federal University Kashere (FUK), Gombe State**,
under the supervision of academic and technical staff.
The motivation is to improve feature sensitivity in steganalysis tasks, where
hidden information introduces subtle statistical deviations in image data.
## Key Idea
Instead of using a fixed activation function throughout the CNN, this approach:
- Observes activation output behavior
- Dynamically selects ELU, ReLU, or LeakyReLU
- Iteratively refines the model’s non-linearity
This hybrid strategy aims to enhance representational robustness for detecting
steganographic patterns.
## How to Run
1. Install Dependencies
```bash
pip install -r requirements.txt
```
2. Run the Model
```
python src/hybrid_lu_cnn.py
```
3. Run Tests
```
python src/hybrid_lu_cnn.py
```
## Results
Experimental visualizations such as activation distributions and architecture
comparisons are stored in the results/ directory.
### Activation Function Analysis & Impact
The figure above compares **ReLU**, **ELU**, and **Leaky ReLU** activation functions across negative and positive input regions.
**Key observations:**
* **ReLU** performs efficiently for positive inputs but completely suppresses negative values, which may lead to information loss (the *dying ReLU* problem).
* **ELU** preserves negative information smoothly, improving gradient flow and stability during training.
* **Leaky ReLU** maintains a small gradient for negative inputs, reducing neuron inactivity while remaining computationally simple.
**Hybrid Impact:**
By allowing these activation functions to **work together dynamically**, the hybrid LU strategy combines:
* ReLU’s efficiency …