Deepfakes Now Outsmarting Detection By Mimicking Heartbeats

BERLIN — Digital doppelgängers just developed a pulse. Researchers hunting deepfakes long relied on a seemingly foolproof detection method: AI couldn’t fake the subtle skin color changes caused by your pulse. That certainty has now collapsed, as scientists discover modern deepfakes inadvertently preserve the heartbeat patterns from their source videos, making our most trusted detection tools suddenly unreliable.

For years, cybersecurity experts thought they had deepfakes all figured out. Since AI wasn’t explicitly programmed to mimic subtle color changes in human skin caused by blood flow, researchers believed fake videos would lack these physiological signals. But according to a new international study published in the journal Frontiers in Imaging, that defense has now crumbled.

Previous research has suggested that deepfake creation processes erased the subtle heartbeat-related signals in videos, making them a reliable tool for detection. However, this study challenges this idea, showing that modern deepfake methods no longer remove these signals.

This revelation means many existing deepfake detection tools may be less effective than previously thought. Banking on pulse detection as a shield against deepfakes may no longer be an option.

How Fake Videos Acquired a Pulse

How can you detect a heartbeat in a video? Our heartbeats cause subtle color variations in our skin as blood flows through vessels near the surface. This phenomenon, invisible to the naked eye, can be detected using a technique called remote photoplethysmography (rPPG).

Think of rPPG as a contactless way to measure someone’s pulse through video alone. Special algorithms analyze the almost imperceptible color changes in facial skin to extract heart rate information. Until now, this technique has been considered one of the most promising methods for identifying deepfakes.

For their study, the researchers created a pipeline specifically designed to extract these physiological signals from videos. They analyzed both genuine videos and deepfakes, focusing on heart-beat-related signals that typically occur between 0.7 Hz and 3 Hz (equivalent to 42-180 beats per minute).

The team recorded videos of twelve diverse participants under controlled conditions with uniform lighting. During each session, participants performed various activities like talking, reading, and interacting with the recording supervisor. Crucially, the researchers also measured the actual heart rates of selected participants using electrocardiogram (ECG) and photoplethysmography (PPG) sensors to establish ground truth data.

Deepfake Evolution

Using these recordings, the researchers generated their own deepfakes using multiple methods. One method used a special system that combines two decoders to create realistic deepfake videos, resulting in 858 videos of different people and 156 original, unedited videos. They also created additional deepfakes using an open-source tool called DeepFaceLive, which performs real-time face swapping.

The generated deepfakes were impressively high-quality, avoiding common mistakes that plagued earlier generations of fake videos. For additional testing, the team incorporated deepfakes from public datasets, including the Korean DeepFake Detection Dataset (KoDF).

When the researchers analyzed the deepfakes, they discovered that the fake videos contained valid heart rate signals that closely matched those in the original source videos. The correlation was particularly strong in the DeepFaceLive-generated fakes.

This means that deepfake technology is unintentionally preserving physiological signals from source videos, effectively transferring the “heartbeat signature” to the fake videos.

Many researchers had built detection systems on the premise that deepfakes couldn’t replicate these subtle biological signals. Now, those systems may need complete redesigns.

There is one silver lining. While the deepfakes preserved heart rate signals, these signals were generally weaker than in genuine videos. The average signal-to-noise ratio was -1.97 dB for genuine videos compared to -3.35 dB for deepfakes, indicating that the pulse signal in deepfakes, while present, is of lower quality.

The Next Frontier in Deepfake Detection

What does this mean for our ability to identify fake videos? According to the researchers, we need to shift away from simply looking for the presence or absence of heart rate signals. Instead, future detection methods should analyze how these signals are distributed across the face.

It’s not just about whether a pulse exists in the video, but whether that pulse behaves naturally across different facial regions. Genuine videos show specific patterns of blood flow that deepfakes might not perfectly replicate, despite capturing the overall heart rate.

Deepfakes have evolved from obvious fakes with visible artifacts to sophisticated manipulations that can fool human observers. Political figures like Barack Obama, Donald Trump, and others have already been targeted, raising concerns about the potential for misinformation campaigns using fake videos.

The race between creators and detection technology is on. Deepfakes are becoming more sophisticated than we anticipated, and our detection methods must evolve just as quickly as the deepfakes themselves.