Phys-EdiGAN: A privacy-preserving method for editing physiological signals in facial videos

Xiaoguang Tu, Zhiyi Niu, Juhang Yin, Yanyan Zhang, Ming Yang, Lin Wei, Yu Wang, Zhaoxin Fan, Jian Zhao

Abstract: Highlights•This paper presents Phys-EdiGAN, a novel method for editing physiological signals in facial videos using Conditional GAN. The method enables precise customization of physiological privacy information, effectively preventing privacy leaks while preserving high video quality and ensuring that facial identity remains consistent.•We propose a Self-Critic Module to learn the latent mapping between rPPG signals and video frames, guiding the generator to embed target rPPG signals into facial videos. An Identity Preservation Component is added to maintain the person’s identity in the generated videos.•Our experiments on databases such as UBFC-rPPG and VIPL-HR demonstrate that Phys-EdiGAN can generate more natural and realistic modifications of physiological signals, maintain overall video consistency and quality, and is highly adaptable, allowing for a range of personalized applications.