Ukrainian Catholic University

Faculty of Applied Sciences

Data Science Master Programme

Responsible Data Science final project

Authors:

Anastasiia Khaburska

Anton Shcherbyna

Vadym Korshunov

Yaroslava Lochman

Interpretability of Fake Face Discrimination

The project idea was inspired by Kaggle competition "Real and Fake Face Detection" organised by Department of Computer Science, Yonsei University.

In this work, we took StyleGAN - pretrained generator of fake faces - and StarGAN - pretrained discriminator of faces - and investigated behaviour of this discriminative CNN model for the Transparency and Interpretability goals. We interpreted the model's output on the fake generated images. To accomplish this task, we explored the most popular feature visualisation techniques and tried a few of them implemented in Pytorch:

• Vanilla Backpropagation (with Smooth Gradient)
• Guided Backpropagation
• Gradient-weighted class activation mapping (GradCAM)

📓 This Colab Notebook is an interactive Report on our project.

All the code, we where working on, may be seen our GitHub repository

Also, a great help for us in understanding and implementing of CNN visualisation techniques was this GitHub repository with PyTorch CNN Visualizations