Cover photo by ksk_1990
Once considered a reliable method to detect deepfakes, 'heartbeat detection' is now challenged by new research. Modern, high-precision deepfakes can replicate even subtle physiological signals like human heartbeats.
This article provides a clear overview of the technical background and future directions for detection methods.
Visual Realism Brings 'Heartbeat'
Recent deepfake technology has advanced to not only mimic visual realism but also physiological characteristics. 'Heartbeat visualization' is particularly noteworthy. Detection methods using rPPG (remote photoplethysmography) once deemed a heartbeat as proof of authenticity.
Photo by take_apii1
However, recent studies show that videos generated using deep learning can 'accidentally' inherit subtle blood flow patterns from the original footage, even replicating fake heartbeats.
Risk of False Detection and Technical Challenges
Research at Humboldt University found that detection models extracting heartbeats from about 10 seconds of facial footage often detected 'heartbeats' in fake videos, mistaking them for real. This undermines the previous assumption that 'heartbeat equals authenticity.'
The persistence of subtle facial tones and movements during video generation can appear as 'heartbeats.' Researchers are now shifting towards analyzing local blood flow patterns and temporal changes in faces. They emphasize the need to combine multiple detection technologies rather than relying on a single indicator.
Warnings from Advancements in AI Technology
This research holds significant meaning for us creators, as it prompts us to reconsider what 'real' truly means.
Photo by mizuno takao
While integrating AI into creative work, how do we maintain reliability and originality? Additionally, creating works themed around deepfakes can spark social discussions.
