PPHAR-GAN: Privacy-Preserving Action Recognition Using Generative Adversarial Network-Based Deep Learning In Surveillance Videos

Abstract

Maintaining public safety and security, particularly in urban areas, now depends on the ability to recognize human behavior from surveillance images. Surveillance cameras in public spaces are becoming increasingly common for the same reason. Video analytics and human activity recognition have been simplified by artificial intelligence (AI). Advances in deep learning and generative adversarial network (GAN) architectures may play a major role in improving research on action recognition. However, preserving the identity of the individuals in the video is necessary for successful identification of human activities. There are instances where safeguarding privacy is essential to upholding the public interest. This means that certain privacy preservation-related limitations exist in the current research. In order to address this issue, we presented in this study a unique GAN architecture called privacy-preserving human action recognition GAN (PPHAR-GAN). This design takes use of effective human action recognition and identity concealment. Our suggestion was for an algorithm called Privacy-Preserving Human Action Recognition (PP-HAR), which makes use of PPHAR-GAN to maximize efficiency in human action recognition while maintaining privacy. The suggested technique was tested using the JHMDB and DALY benchmark datasets. The findings showed that, with the greatest accuracy of 98.51%, the suggested algorithm beats many other deep learning-based models already in use. As a result, real-time applications needing privacy-preserving human action recognition can use our framework.