This is a generative adversarial network created for ECE2195, Information Security, at the University of Pittsburgh. Working with a modified version of the Sokoto Coventry Fingerprint Dataset (See Usage.pdf for details).
The premise of the project is to create a neural network which can regularly generate fake fingerprints which can fool a neural network designed to classify fingerprints as a means of biometric authentication.
The code was modified from this example tutorial on the PyTorch website. The SOCOF dataset was modified to fit the fixed input size (64x64 pixel RGB images).
The software depends on a number of Python modules to do its job, centered around the PyTorch suite. Besides Python 3, a few prerequisites are required, installed via pip
pip3 install numpy matplotlib
pip3 install torch torchvisionIncluded in the repository is the state of a pretrained generator such that the demo program can be run immediately. There is no need to run the training program unless you wish to modify it. Simply follow the demonstration instructions below.
Training the dataset takes a few hours on a laptop, as we did not have a GPU at our disposal to exploit. The original code includes additional setup code to offload training to a GPU.
In the data folder, two folders store the preprocessed training data. The positive directory contains all the left-hand fingerprints of the SOCOF dataset, and was used to represent known fingerprints. The negative directory contains all the right-hand fingerprints, and may also be used to train.
To train a network and output intermediate images, run main.py:
python3 main.pyEvery 10 generations, a sample of the generated fingerprints is saved to the working directory. On completion of training (25 epochs), a plot of loss versus generation will be presented, and the internal states of the networks will be saved to the working directory.
The demo.py program creates a new discriminator network representing the biometric authorization system which is trained on the original dataset. Then, a new generator network is created which uses the weights and paramters captured during the training process. After setup, the generator will produce new fingerprints from a random input vector, display them, and display the output from the discriminator for that fingerprint.
The demo discriminator only uses the original dataset for training, representing a group of known fingerprints for authorized users. The demo generator takes as input a random vector and produces several batches of new fingerprints, which are then checked against the discriminator.
After running the demo program, any generated images which had confidence > 90% will be saved to the working directory for inspection. Run the demo after training:
python3 demo.py