README.md

LockMe

Welcome to my software engineering class final project.

The project consists of a file encryption app. Login is via face detection, thus make sure you have a camera connected to your PC.

Usage

Installation

To activate the system follow these steps:

1. run: pip install cmake
2. then run: pip install -r requirements.txt
3. use a linux CMD to rebuild the model from parts in the model/model_parts directory using the command:cat model.part* > model.pth
4. make sure the model name is written with the correct path in the configurations file in the model's directory
5. you are then ready to run: python main.py

running example

Terminal UI view:

Tkinter based GUI view:

Dataset:

• The dataset is a combination of samples I transformed from AT&T, LFW and my own images.
• I used 3 functions to augment the images:
  1. horizontal flip
  2. brightening
  3. darkening
• To recreate a similar dataset use the functions in the model/dataset.py file.

Labels definition:

• 1 - for samples of different subjects
• 0 - for samples of the same subject

examples:

1. image pairs: labels: [0. 1. 0. 0. 1. 0. 0. 1.]

2. image pairs: labels: [0. 0. 0. 1. 0. 0. 1. 1.]

The Model

• I used the classic SNN architecture using binary cross entropy loss.
• I tried a few models and loss functions such as triplet loss and contrastive loss with different learning rates but BCE gave me the best results with the amount of data that I had.
• The model is trained for 50 epochs and learning rate of 0.0006 using batch size 128.
• The final result is 80% accuracy.

WIP ans further work:

• Training on CelebA dataset (use complex pairs)
• Try architecture: InceptionResNet-V2
• Try triplet loss