Dutch forensic experts unveil breakthrough heartbeat test to detect deepfakes
Dutch researchers have developed a promising new method for detecting deepfake videos by analyzing subtle facial changes linked to a person's heartbeat—a technique that could help forensic investigators identify manipulated footage in an era of fast-advancing artificial intelligence, according to the Netherlands Forensic Institute (NFI).
Zeno Geradts, a digital forensics expert at the NFI and professor of Forensic Data Science at the University of Amsterdam (UvA), unveiled the method this week at the European Academy of Forensic Science (EAFS) 2025 conference in Dublin, which runs from May 26 to May 30. His technique focuses on detecting blood flow patterns in the face—specifically, the expansion and contraction of small facial veins in sync with the heartbeat.
“As far as I know, this is not yet used in forensic research,” Geradts said in a press statement from the Netherlands Forensic Institute. “We are still in the process of scientific validation, but it is a promising addition to existing methods.”
The new approach uses high-resolution video to detect subtle shifts in skin color caused by blood flow beneath the surface—changes that occur with each heartbeat. These minute shifts are especially visible around the eyes, forehead, and jaw, where blood vessels lie just below the skin. According to Geradts, this biological signal is absent in deepfake videos, making it a key indicator of video authenticity.
Geradts’s idea to use heart-rate detection in digital forensics originated over a decade ago. Around 2012, the NFI was occasionally asked to analyze so-called snuff films—videos depicting extreme violence, including torture and murder, which sometimes circulate via dark web platforms or encrypted messaging apps.
While working on these cases, Geradts came across a Massachusetts Institute of Technology (MIT) study indicating that facial veins could be used to monitor heart rate. “I immediately realized that we could use this for image detection,” he said. “We had something valuable in our hands.”
At the time, the idea couldn’t be applied due to poor image compression technology. “Large video files are reduced in size, and during that compression, the color differences per heartbeat were lost,” Geradts explained in comments released by the Netherlands Forensic Institute. “But now, thirteen years later, compression methods have improved, and we can detect even the slightest discoloration caused by pulsing blood flow.”
The new research was initiated by Paula Pronk, a scientific associate at the NFI, and later expanded by intern Sanne de Wit. De Wit, who holds a bachelor's degree in Computer Science from TU Eindhoven and is completing a master's in Forensic Science at the UvA, gathered new data and trained the model. She filmed test subjects wearing both a smartwatch and a heart monitor, then compared those readings to subtle skin tone changes visible on camera.
De Wit analyzed 79 facial points where color changes caused by the heartbeat could be reliably measured. She conducted tests under three conditions: with the subject sitting still, during significant movement, and in low lighting. The study confirmed a strong correlation between visible color changes and the measured heart rate in all settings. Literature also shows that the method works across all skin tones. De Wit is now expanding the dataset and further training the detection model.
Deepfake videos—AI-generated content in which real people appear to say or do things they never did—pose growing risks to individuals and society. Victims may be inserted into explicit content or misrepresented in fake news videos, and geopolitical tensions heighten the stakes for such manipulations. As the technology behind deepfakes advances, distinguishing real footage from fake becomes increasingly difficult with the naked eye. “Sometimes I worry that soon, no one will believe in real images anymore—that everything will be seen as fake,” Geradts said. “What will still be true then?”
Although the blood flow method is not yet ready for courtroom use, it adds to a growing toolbox of detection techniques developed by Geradts and his team. Previous methods include Electric Network Frequency (ENF), which detects the flicker pattern of electrical lighting in a video to determine when it was filmed, and Photo Response Non-Uniformity (PRNU), which identifies the unique "fingerprint" of a camera based on how its sensor reacts to light.
More traditional forensic techniques also remain in use, such as observing speech abnormalities, irregular blinking, or poorly blended face edges in videos where "face swaps" have occurred. However, newer AI models can generate entire faces without these obvious seams. Chain-of-evidence analysis and commercial AI-detection algorithms—new ones appear every month—further bolster investigations. “Our strength lies in combining and continuously improving all these methods,” Geradts said in the press release.
In addition to his presentation, Geradts is leading a workshop on AI detection at the Dublin conference. Participants are asked to judge whether sample videos are real or fake and are given tools to create deepfakes themselves. The workshop also teaches detection techniques using current tools.
Geradts has led similar sessions at international events, including the 2024 conference of the American Academy of Forensic Science (AAFS). However, he notes that no workshop is ever the same, given the rapid pace of AI development.
“I could give a new workshop every month,” he said in an NFI statement. “The developments are going so fast. But I haven’t yet seen a deepfake where the facial heartbeat is visible. Could that change after we publish our findings? Maybe. But this method should be reliable for at least the next two years. In this field, you don’t build tools for eternity.”
