Abstract:
An individual’s face can provide important insight about his personal traits like age, psychology, health, ethnicity, emotions, kinship, and much more. The biometric potentials of facial images make them an ideal tool for various forensic inferences. One such interesting area of research is the detection of criminal tendencies in people from their facial images. Several studies have proposed machine and deep learning based solutions for this purpose. However, none have discussed the impact of various demographics on the performance of such systems. In this thesis, we provide an indepth analysis to measure the impact of demographics i.e. age, gender, and ethnicity on facial image-based criminality detection systems. For this purpose, we have prepared a balanced dataset as there was no such dataset available with required age, race, and gender splits. We have developed a system based on pretrained convolutional neural network architecture (FaceNet) and evaluated its performance in perceiving criminal tendencies from facial images under the impact of different demographic groups. In measuring the impact of gender on classification output, we achieved 82.7% accuracy in classifying images of male subjects while in the case of female subjects our model gives 87.4% accuracy. We then measured the impact of age in which we considered three age groups and achieved 80.8%, 84.2%, and 80.2% accuracies in the 15-30, 31-45, and 46-60 age groups respectively. Finally, we measured the impact of ethnicity on classification output and achieved 77.4%, 84.2%, and 82% accuracies in African, Hispanic, and White ethnicities respectively. The analysis presented in this study can prove vital for the development of robust and unbiased systems that can provide reliable proactive solutions for the security of all communities.
