Deepfakes Now Replicate Heartbeats, Increasing Detection Difficulty

Picture a world where deepfakes are so advanced that no tool can expose them. This would be a goldmine for criminals and hostile governments, who could exploit them to discredit political opponents or falsely accuse human rights advocates.

This alarming scenario hasn’t come true yet, but for years, deepfake creation techniques have been in a constant “tech arms race” with detection tools. Now, researchers have revealed that deepfakes have taken a major leap forward—missing pulse signals can no longer expose them.

“For the first time, we demonstrate that recent high-quality deepfakes can include a realistic heartbeat and subtle facial color shifts, making them significantly more difficult to spot,” said Dr. Peter Eisert, a professor at Humboldt University of Berlin and lead author of a new study in Frontiers in Imaging.

Deepfake creators use deep learning to modify videos and audio, changing facial expressions and gestures by swapping them between different individuals. While people can use this for harmful purposes, it’s not always malicious—apps that transform users into cats or age them digitally have become hugely popular as harmless entertainment.

The study of light transmission through the skin and underlying blood vessels has long been crucial in medicine, such as in pulse oximeters. Its digital counterpart, remote photoplethysmography (rPPP), is an emerging telehealth tool that uses webcams to estimate vital signs. In theory, rPPP could also be applied in deepfake detection.

In recent years, experimental rPPP-based deepfake detectors have shown strong ability in differentiating between real and deepfake videos. These successes led some experts to believe that current deepfakes couldn’t yet replicate a realistic heart rate. However, it now seems that this optimistic view is no longer accurate.

Pretend until you succeed

Eisert and his team developed an advanced deepfake detection system that automatically extracts and analyzes the pulse rate from videos. It employs innovative techniques to account for movement and eliminate noise, requiring only a 10-second video of a single person’s face for operation.

The authors also compiled their own dataset of driving videos, using them to generate deepfakes with various target identities while preserving the facial movements from the original footage.

During the filming process, an ECG monitored the protagonists’ heartbeats, enabling the researchers to verify that the rPPP measurements from their new detector were highly precise.

The estimates of the pulse rate differed by just two to three beats per minute from the actual rate. To further test their detector, the authors also applied it to two older, well-known video datasets of real people. In these cases as well, they successfully extracted heartbeat signals from all the authentic videos.

But what would occur if they applied the same detector to analyze known deepfakes?

To test this, Eisert and his team employed recent deepfake techniques to swap faces in genuine videos from their collection. To their surprise, their detector detected a pulse in the deepfakes as well, even though they hadn’t intentionally included one. This artificial pulse often appeared strikingly realistic.

Finding hope

Eisert stated, “Our findings show that an attacker can add a realistic heartbeat or unintentionally inherit it from the authentic video. Subtle skin tone changes and facial movements from the real person are transferred to the deepfake, replicating the original pulse.”

The authors concluded optimistically, suggesting that deepfake detectors could stay ahead by focusing on detecting local blood flow in the face instead of just the overall pulse rate.

Eisert stated, “Our experiments have demonstrated that while current deepfakes may display a realistic heartbeat, they fail to exhibit physiologically accurate variations in blood flow across both space and time within the face.”

“We propose that the next generation of deepfake detectors should take advantage of this vulnerability in current state-of-the-art deepfakes.”

Read the original article on: Techxplore

Read more: Injectable Pacemaker Regulates Heartbeat for Five Days, Then Dissolves