Noise Generation GAN Based Identity Privacy Protection for Smart City

The development of Internet of Things (IoT) infrastructure in the city leads to the emergence of the concept of smart city, an integrated solution to provide convenience for various applications in our daily life by understanding and analyzing the collected data from multi-sources. However, the collection of facial images collected from various IoT devices such as surveillance cameras, wearable, and mobile devices increases the risk of an individual’s privacy leak. The facial recognition models augment this risk. These models retrieve facial data collected from IoT devices stored in smart city databases to get personal identity information. With extensive utilization of such IoT devices, which serve as a visual data collector, we compromise the person’s identity. Therefore, to protect the privacy of image data from a database, we propose a Sensitivity Map Noise-Adding model based on generative adversarial networks to provide privacy for facial images against the malicious use of the face recognition models. The proposed models work as a black-box model that does not require any architectural information or the parameters of the target model. Additionally, the model runs at a real-time speed and the average run time for one operation is less than 12 milliseconds. The protection can be deployed for both local images and streaming videos. The data privacy protection is based on our proposed concept of the Sensitivity Maps, which summarizes the effectiveness and efficiency of adding noises on each pixel on the original image to interfere with the target model’s performance. We have built a new dataset of facial images containing 102 celebrities for the proposed model to be trained and evaluated. The experimental results prove the advantage of the proposed method against protecting the identity information in facial images.