Examining Gender Bias of Convolutional Neural Networks via Facial Recognition

Abstract

Image recognition technology systems have existed in the realm of computer security since nearly the inception of electronics, and have seen vast improvements in recent years. Currently implemented facial detection systems regularly achieve accuracy rates close to 100 percent. This includes even challenging environments, such as with low light or skewed images. Despite this near perfect performance, the problem of gender bias with respect to accuracy is still inherent in many current facial recognition algorithms. This bias needs to be addressed to make facial recognition a more complete and useful system. In particular, current image recognition system tend to have poor accuracy concerning underrepresented groups, including minorities and female individuals. The goal of this research is to increase the awareness of this bias issue, as well as to create a new model for image recognition that is gender independent. To achieve this goal, a variety of Convolutional Neural Networks (CNNs) will be tested for accuracy as it pertains to gender bias. In the future, the most accurate CNNs will then be implemented into a new network with the goal of creating a program which is better able to distinguish individuals with a high accuracy, but without gender bias. At present, our research has identified two specific CNNs, VGG-16 and ResNet50, which we believe will be ideal for the creation of this new CNN algorithm.