Mitigating Racial Biasness in Facial Recognition Technology Using Generative Models

In recent years, several researchers have worked to lessen biasness in face recognition systems since it had produced numerous problems, from people getting falsely accused of minor offences up to murder, getting wrongfully arrested and in some extreme cases, getting killed, and these happened particularly for people of colour. This thesis aims to contribute to the data science field by investigating and implementing generative models as a potential assistance to expand the diversity of datasets used in facial recognition technologies and help mitigate racial biasness in facial recognition systems. The CelebA Dataset, containing more than 100,000 unique photos, was utilized to train our StyleGAN 2 model, to generate synthetic realistic images and the FairFace dataset which has a diverse dataset over 100k images of both males and females for our recognition model. East Asian, African, Caucasian, Indian, Middle Eastern, Latino, and Southeast Asian are the racial categories that have been identified. We used InceptionResnetV1 for feature extraction, MTCNN for face detection and then ran our recognition Model on our diverse dataset with and without synthetic images. After using our own generated synthethic data, we saw accuracy gains for a few of the races, including East Asian, African, Indian, Middle Eastern, and Latino, which demonstrated that the accuracy level increased by more than 10% in some cases.