Combining Deep Privacy with an Attribute-driven Generative Adversarial Network to Preserve Gender and Age in De-identified CCTV Footage

Abstract

A surveillance camera is an efficient solution to prohibit crimes for both small and big businesses, and is broadly utilized in big cities. Today, the police force can only access the camera footage for further investigation after an act of crime. In order to observe, find patterns, and react appropriately to an event, the Oslo Police wants to use its own CCTV cameras and analyze such footage in real-time. To investigate real-time CCTV footage and share such footage with a third-party for analyzing, the people in the footage need to be de-identified. In this thesis, we focus on de-identification of CCTV footage, preserving age and gender for more precise context information. DeepPrivacy is a neural network model that creates new faces using image in painting. It is found to be suitable for de-identification of CCTV footage but the creators did not intend to preserve age and gender. The thesis proposes combining DeepPrivacy and an attribute-driven network to enforce preservation of age and gender, and performs experiments on two state-of-the-art, attribute-driven Generative Adversarial Networks (GANs),AttGAN, and StarGAN v1. These networks are designed to keep the input image intact while changing specific attributes. The thesis also studies the option of changing the subjects’ skin tone to a specific color to bypass potential ethnicity bias.