The race between deepfake creators and detectors has entered a new phase, with researchers from the Netherlands revealing tech that reads the human heartbeat through video analysis. While artificial intelligence (AI) generates increasingly convincing deepfake videos, forensic researchers are responding by building an ever-expanding toolkit of detection methods.
Detecting deepfakes requires multiple approaches working in concert – from analysing camera fingerprints and electrical grid frequencies, to examining speech patterns and facial inconsistencies. No single method provides definitive proof, but the combination of classical inspection techniques and cutting-edge AI analysis creates a robust defence against synthetic media.
Now, researchers at the Netherlands Forensic Institute (NFI) have developed a method that adds biological authentication to this arsenal, detecting deepfakes by analysing blood flow patterns in human faces – patterns that current deepfake generation tools cannot yet replicate.
Zeno Geradts, a forensic digital researcher at the NFI and one of the directors of the AI for Forensics Lab, which is jointly operated with the University of Amsterdam, recently presented this groundbreaking research at the European Academy of Forensic Science conference in Dublin.
“If you have good-quality images, there are small arteries in your face that expand slightly with each heartbeat,” he said.” That colour difference per heartbeat can be seen in a video. This is also a method used in the medical industry and smartwatches, but our method works remotely.”
The technique builds on existing remote photoplethysmography (RPPG) technology, which measures pulse rates through subtle colour changes in skin caused by blood circulation. When applied to video analysis, researchers can detect the rhythmic colour variations around the eyes, forehead and jaw where blood vessels lie close to the skin surface.
From snuff films to deepfake detection
The origin of this research traces back to 2012, when the NFI was occasionally asked to investigate so-called snuff movies – extremely violent films circulating on clandestine channels. Geradts needed to determine whether people in certain films had actually died.
“I accidentally came across a publication from MIT where researchers had discovered that you can measure someone’s heartbeat based on the small arteries in the face,” recalled Geradts. “I immediately knew we could use this for image detection.”
However, the technology wasn’t ready at the time. Video compression techniques destroyed the subtle colour differences that indicate heartbeat. Thirteen years later, improved compression methods preserve enough image quality to make the detection possible.
A digital research team of the NFI analysed 79 facial points where colour differences per heartbeat could be measured, testing under various conditions, including movement and low light. Results showed consistent correlations between measured heartbeat (via smartwatch and heart rate monitor) and facial colour variations under all circumstances. Literature indicates the method works across all skin tones, although darker skin presents additional challenges due to reduced colour contrast.
Expanding the detection arsenal
Blood flow detection represents the latest addition to a comprehensive forensic toolkit. The NFI employs multiple detection methods simultaneously, each suited to different scenarios and video qualities. The power lies not in any single technique, but in their combined application.
Electric Network Frequency (ENF) analysis examines the subtle flickering of lights in videos caused by variations in power grid frequency, which helps determine when the footage was recorded. Photo Response Non-Uniformity (PRNU) creates a “fingerprint” of specific cameras by analysing how individual pixels respond to identical light levels.
Traditional inspection methods remain valuable, including the detection of speech anomalies, irregular blinking or poorly rendered facial edges. AI detection algorithms trained on authentic versus fake content can sometimes identify the specific generation tool used to create a particular deepfake.
“The strength of good deepfake detection lies in using a combination of classical methods and AI,” said Geradts. “You have to look at the complete picture – both image and audio. Synthetic voices are challenging to make realistically.”
Each method contributes crucial evidence, but none stands alone; instead, they complement one another. High-quality source material might favour blood flow detection, while compressed YouTube videos might require different approaches. The forensic value emerges from cross-validation across multiple techniques.
The perpetual arms race
The effectiveness of detection methods faces constant challenges as generation technology improves. Conference feedback suggested that future deepfake training might incorporate heartbeat signals, potentially neutralising this detection method.
“It remains a cat-and-mouse game,” said Geradts. “We’ll have to keep researching and discovering new methods to stay ahead of criminals.”
State actors with substantial resources and knowledge of detection methods pose particular challenges. However, major tech companies that develop legitimate deepfake tools often provide their own detection capabilities alongside their generation software.
The research paper on blood flow detection is nearing completion and is expected to be published within six months. Until scientific validation is complete, the method cannot be used in forensic investigations; however, Geradts expects deployment for suitable cases with high-quality source material.
Implications for digital evidence
The development highlights growing concerns about the proliferation of deepfakes across various sectors. From fraudulent business communications requesting financial transfers to non-consensual intimate imagery, the potential for harm extends far beyond entertainment applications.
For investigators and legal systems, the challenge intensifies with each technological advancement. Forensic analysis, which once took weeks, may soon require months as detection methods become more sophisticated and comprehensive.
“Sometimes I worry that eventually nobody will believe real images anymore,” said Geradts. “That everything will be seen as fake. What is still true then?”
The NFI’s multi-method approach provides a template for other forensic institutes facing similar challenges. As deepfake technology democratises and its quality improves, robust detection capabilities become essential infrastructure for maintaining digital evidence integrity in legal proceedings.
The heartbeat detection method may face obsolescence as generation tools evolve, but it demonstrates the ongoing innovation required to preserve trust in digital media. For European law enforcement and judiciary systems, such research represents a crucial investment in future investigative capabilities.
