Semi Adversarial Network

Neural Network exam's project. Keras implementation of the work done in:

Semi-Adversarial Networks: Convolutional Autoencoders for Imparting Privacy to Face Images by V. Mirjalili, S. Raschka, A. M. Namboodiri, and A. Ross. https://arxiv.org/abs/1712.00321

Our work is based on the preprocessed Celeb-A dataset provided by iPRoBe-lab's implementation

For more details about this project read our final report.

Usage

Dataset

First you need to download and preprocess dataset.

CelebA

Our work is based on the Celeb-A dataset provided by iPRoBe-lab's implementation

• Download training set: here
• Download test set: here

Place both images-dpmcrop-train and images-dpmcrop-test folders in repo's root and just run

cd src/pre_processing
python prepare_dataset_celeba.py

Notes

At the end of pre-processing the dataset folder will be structured as follows:

dataset
|
+-- celebA
|   |   
|   +-- test
|   |   +-- female
|   |   +-- male
|   +-- train
|   |   +-- female
|   |   +-- male
|   +-- validation
|   |   +-- female
|   |   +-- male
+-- prototype

---

Pre Training

In this phase models of the Semi-Adversarial-Network are trained independently. AutoEncoder and Gender-Predictor have been pre-trained by us.

cd src/

For pretraining the Autoencoder

python pretrain_autencoder.py

For pretraining the Gender Classifier

python pretrain_genderclassifier.py

Face-Matcher uses downloaded weights (in the same way of the reference paper) during Further-Training so we didn't need to pre-train it.

Obtained weights can be found here: https://drive.google.com/file/d/1OgnvG74Qru9tbeXImUldcr60HuJoLrov/view?usp=sharing

---

Further Training

Train the autoencoder using other modules' feedbacks.

The complete Semi-Adversarial model (modules/san.py) is formed by the following NNs:

• AutoEncoder
• Gender Classifier
• Face Matcher

Face Matcher and Gender Classifier are not trainable components during this phase.
Be sure to have placed pretrained weights in /weights and named them as follows:

- autoencoder_pretrain_weights.h5
- genderclassifier_pretrain_weights.h5
- facematcher_pretrain_weights.h5

These are the only weights considered by further_train.py, so if you retrain one of the modules, please rename obtained weights accordingly or manage weights loading as you prefer in each modules in /src/modules.
Then you can run further train with

cd src/
python further_train.py  

Obtained weights will be saved in /weights with a timestamp attached to the file name.
Obtained weights can be found here: https://drive.google.com/file/d/1O4hjK8BBsD922SVnFEkPDitE4MEZ5ss8/view?usp=sharing

This project was developed with passion by: Bruno Marino & Gianluca Pepe.