So I’ve released my paper on Facial Recognition I did for my Digital Forensics course. This course was taught by Det. Watson with the Reno PD, and used a lot of state of the art forensic tools. It is part of the UNR cybersecurity push that has been great to be a part of.
I did a paper on Data Augmentation. It’s called Data Augmentation: Demographics and Race in Facial Recognition and the link is here: https://docs.google.com/document/d/1oYfU4IIEaKcvWiExSJFtaT7eyl-k76POWD82bI69h9I/edit?usp=sharing
Basically, I used a Generative Adversarial Network to create faces from a Tensorflow Hub pretrained GAN. Then I interpolated between these faces as vectors on a tensor, creating blends of faces from the same demographic categories. Two example of this are seen here (Bruno Mars and George Eads):
Each face is increasingly dissimilar for the first face. By attempting to match the celebrity face at the top, we can begin to understand disparities between demographics in facial recognition, a topic which has been called into question by ANSI, the American National Standards Institute. Each face down the interpolation is less like the actual GAN generated celebrity, and should define a boundary within the interpolation space, which represents precision of the adversarial images. Ideally, a demographically fair model will draw the same boundaries across these distortions regardless of demographic and between demographic classes.
The demographic categories for which data was generated were: Female and Male, African and Caucation with four total categories, although this could be expanded with the code included.
All of the data used in this paper is available in the Github Repository here:
https://github.com/davidgabriel42/Dataset-CS704-Data_Augmentation-Demographics-Race