Autoencrypt

Author: Ethan Robinson Started: November 16, 2023

Overview

Autoencrypt is a machine-learning-based image reconstruction tool that reconstructs specific images based on a unique input sequence, which functions similarly to a "password." Rather than encryption or decryption, this project focuses on using trained neural networks to produce an accurate reconstruction of an image only when provided with the correct input sequence.

Table of Contents

Overview Features Installation Usage Project Structure Implementation Details

Features

Input-Specific Image Reconstruction: The neural network reconstructs an image only when the correct input sequence is provided. Zero-Padded Centered Images: Trains models to handle smaller images, reconstructing them as centered on a black background to accommodate dimension constraints. Noise Generation for Incorrect Inputs: The model is trained to output random noise for all input sequences that do not match the specific image-generating input. Customizable Neural Network Architecture: Flexible and modifiable for experimentation with different model structures. Planned Enhancements Zero-Padded Reconstruction: For smaller images, the model will reconstruct them as centered on black backgrounds using zero-padding. This adjustment accommodates the model’s output dimension requirements for generating consistently large images. Noise Generation for Incorrect Inputs: The model will be trained to generate random noise for any input sequences other than the specific “password” input, adding a layer of security and unpredictability.

Installation

Clone the Repository:

git clone https://github.com/erobinson-1997/Autoencrypt.git
cd Autoencrypt

Install Dependencies:

Make sure you have Python installed, then install required packages:

pip install -r requirements.txt

Usage:

To run the main script, provide a password and an image file path to initiate the reconstruction process.

Workflow

1. Training the Model:

   • The model is trained to reconstruct an image based on a specific input sequence (password).
   • After training, the model is saved to the file system for future use.
   • The user can confirm the reconstruction quality by comparing the reconstructed image with the original.

2. Reconstructing an Image:

   • The user selects a saved model from the file system and provides the correct input sequence (password).
   • The model attempts to reconstruct the image based on the input sequence.

3. Secure Workflow:

   • Once satisfied with the reconstruction quality, the user can delete the original image.
   • The model and input sequence together can be used to regenerate the image later.

Command-Line Usage

Train the Model

To train the model and save it to the file system:

python main.py train <password> <path/to/image_file.jpg> <model_name>

<password>: The sequence required to train the model. <path/to/image_file.jpg>: Path to the image to be reconstructed. <model_name>: Name of the model file to save (e.g., model_1.h5).

Example:

python main.py train mySecretPassword image.jpg my_model

Reconstruct an Image:

To reconstruct an image using a saved model:

python main.py reconstruct <password> <model_name>

<password>: The correct input sequence for the model. <model_name>: Name of the saved model file.

Example

python main.py reconstruct mySecretPassword my_model

Project Structure

main.py: Entry point for running the reconstruction process. services/password_processor.py: Converts the password string into a format suitable for model input. services/image_processor.py: Preprocesses images to prepare them for training or reconstruction. models/ImageDecoder.py: Defines the neural network architecture for reconstructing images based on input sequences. README.md: Documentation and instructions for the project. requirements.txt: Lists the necessary packages for the project.

Implementation Details

• Image Reconstruction: The ImageDecoder model uses layers such as LSTMs and Conv2D Transpose to upscale and reshape encoded data back into image form.

• Zero-Padding: Smaller images are padded with zeros during training to ensure consistent output dimensions, allowing the model to generate centered reconstructions on a black background.

• Noise for Incorrect Inputs: For any incorrect "password" sequence, the model produces a random noise image, reducing unintended access and enhancing security.