DIFFUSE Disentanglement of Features For Utilization in Systematic Evaluation

Purpose and goal:

We explore an approach for automatically collecting large synthetic datasets of diverse facial images, each with distinct characteristics. The project focuses on investigating the use of disentangled and interpretable latent spaces within state-of-the-art generative frameworks, such as Generative Adversarial Networks and 3DMM, to validate other machine learning models and generate synthetic data for training purposes.

The primary emphasis is on generating facial images, with driver authentication being one of the application areas where these methods is tested. Expected results and effects:

Two methods for feature disentanglement have been tested. One utilizing Basel Face Model and another one with Generative Adversarial Networks. The results have shown that the improvements made in this project to the Basel Face Model allow us to generate data evenly distributed across several attributes such as gender and ethnicity, which can better balance models trained on biased datasets, reducing their bias.

Our new method related to the Generative Adversarial Network also shows promising results in its ability to encode image information into a disentangled latent space. Approach and implementation:

Two methods have been explored within the project in order to both capture direct use and future relevance in industry applications. Each with their own advantages, for example the Basel Face Model is easier to use for commercial purposes as the Generative Adversarial Networks are pretrained on datasets for research purposes and would need to be retrained. On the other hand, Generative Adversarial Networks are among the most realistic models currently.