AI Generated Videos Now Mimic Heartbeats Posing New Detection Challenges

Deepfakes Reach New Level of Realism: Berlin Researchers Demonstrate Heartbeat Imitations

The technology behind deepfakes is rapidly evolving. A research team from the Humboldt University of Berlin and the Fraunhofer Heinrich Hertz Institute HHI has now shown that AI-generated videos can even imitate the human heartbeat. This development significantly complicates the detection of deepfakes and presents new challenges for existing detection methods.

Subtle Facial Changes as the Basis for Pulse Measurement

The human heartbeat causes minimal color changes in the skin that are invisible to the naked eye. However, these changes can be detected using photoplethysmography (PPG). PPG uses light-emitting diodes to measure the changes in blood volume caused by the heartbeat and thus determine the pulse. This technology is used, for example, in smartwatches. A further development, remote photoplethysmography (rPPG), enables contactless pulse measurement via video recordings.

rPPG as a Potential Weapon Against Deepfakes – and its Limitations

rPPG technology has been touted as a promising method for deepfake detection. The assumption: Since AI-generated videos cannot reproduce the subtle color changes in the face caused by the pulse, rPPG could reliably expose deepfakes. The Berlin researchers have now refuted this assumption.

Study Demonstrates the Ability of Deepfakes to Imitate the Heartbeat

In their study, published in the journal "Frontiers in Imaging", the researchers present a specially developed "deepfake pipeline". This consists of a deepfake detector based on rPPG and a deepfake generator trained with videos of twelve subjects. The results show that the generated deepfakes exhibit realistic heart rates and can thus bypass rPPG-based detection.

Training Data as the Key to Imitating the Heartbeat

The researchers suspect that the AI is able to learn the subtle color changes in the skin from the training data and apply them when generating deepfakes. Remarkably, the training videos used were only a few seconds long each. This finding underscores the learning ability of modern AI systems and the associated challenges for deepfake detection.

New Detection Methods Urgently Needed

The study shows that the development of deepfakes is progressing faster than the development of detection methods. The ability to imitate physiological signals such as the heartbeat increases the realism of deepfakes and makes their identification more difficult. The researchers emphasize the need for new, more complex detection methods that go beyond simple pulse detection.

Outlook: Deepfakes and the Future of Digital Media

The increasing difficulty in detecting deepfakes poses a challenge for society. The rapid development of AI technology requires continuous research and innovation in the field of deepfake detection to counter misuse and manipulation. Mindverse, as a German provider of AI solutions, is aware of this challenge and is working on innovative approaches to detect and combat deepfakes.