Remote photoplethysmography: advancing robustness, privacy and security

This thesis investigates how to measure a person's heartbeat and other vital signs using only a video of their face. This method, known as remote photoplethysmography (rPPG), does not require any physical contact or wearable sensors. It offers a more comfortable and accessible alternative to traditional health monitoring, especially in everyday or remote settings. While promising, rPPG faces several challenges. The signal can easily be affected by lighting, movement, and other environmental conditions, which makes it unreliable in real-world use. There are also important concerns about personal privacy and the risk that such systems could be misused or attacked.

To make rPPG more robust, this research develops new artificial intelligence methods, including self-supervised learning and transformer-based deep learning models. These methods can detect subtle heartbeat signals more accurately and even without labeled training data. The thesis also focuses on privacy and security. It proposes a method that removes identifying facial information from videos while keeping the heart signal intact. This helps protect individuals when their data is shared. Additionally, two new datasets are introduced to test how rPPG systems respond to physical disruptions and impersonation attacks. These tests reveal serious weaknesses in current methods and provide analysis and tools for developing more secure methods.

Overall, the work helps make video-based health monitoring more reliable, private, and secure, supporting its safe use in future fields like healthcare, driver monitoring, personal wellness, and security applications.