Deepfakes with heart: AI-generated videos now beat biometric detection

Until now, many deepfake detection systems, such as FakeCatcher or DeepFakeON-Phys, relied on global HR signal absence as a distinguishing marker. These models extracted physiological patterns across the entire face and flagged videos lacking consistent biological rhythms as suspicious. The new study reveals that this strategy is no longer enough.

In the deepfake era, it's becoming increasingly difficult to distinguish truth from fabrication, as synthetic media not only mimics human appearance and speech but now even replicates subtle biological signals like heartbeats and skin color, challenging the reliability of traditional detection methods and raising urgent questions about authenticity, trust, and digital identity.

A study titled “High-Quality Deepfakes Have a Heart!” published in the journal Frontiers in Imaging reveals that many advanced deepfakes now carry convincing heart-rate signals that closely mirror those in genuine footage.

This peer-reviewed study by researchers at the Fraunhofer Heinrich-Hertz-Institute and Humboldt University marks a critical turning point in the arms race between deepfake creation and detection. It shows that modern synthetic videos can inherit remote photoplethysmography (rPPG) signals, subtle color changes due to blood circulation, from the source or “driver” videos used during their creation. These findings challenge the current effectiveness of physiological detection methods and signal the need for a dramatic shift in how deepfakes are assessed for authenticity.

How can deepfakes simulate a heartbeat?

The study centers around rPPG, a technique that extracts pulse-related signals from facial videos by analyzing micro color shifts linked to blood flow. Traditionally, researchers assumed that the deepfake generation process, which involves manipulating facial features and re-rendering images, would disrupt or erase these physiological signals. As a result, the absence of a heart rate was widely viewed as a reliable indicator of a fake.

However, the new research disproves this notion. The authors developed a robust signal extraction pipeline, validated against ECG ground-truth data, and applied it to both genuine and synthetically generated videos. Their system aligned facial landmarks frame-by-frame, compensated for motion, and analyzed global rPPG signals using FFT-based frequency analysis. Crucially, the team tested their pipeline across their own dataset of high-resolution videos, deepfakes generated with DeepFaceLive and other advanced autoencoder techniques, and the widely used KoDF dataset.

In all these cases, they found that deepfakes were not only capable of producing plausible pulse rates but also exhibited significant signal correlation with the original driver videos. This was especially true for deepfakes generated with high-end tools like DeepFaceLive, which achieved Pearson correlation coefficients of up to 0.82 when compared with the source video’s rPPG signal. Even older datasets like KoDF revealed that synthetic content could preserve realistic heart rates when crafted with sufficient care.

Why does this undermine current detection techniques?

Until now, many deepfake detection systems, such as FakeCatcher or DeepFakeON-Phys, relied on global HR signal absence as a distinguishing marker. These models extracted physiological patterns across the entire face and flagged videos lacking consistent biological rhythms as suspicious. The new study reveals that this strategy is no longer enough.

Deepfakes are not just generating random noise; they are often embedding inherited heart rate signals from the source footage. The researchers demonstrated that global rPPG signals in deepfakes aligned closely with the original, meaning simple presence or absence of a pulse is no longer a definitive test of authenticity. In fact, the study recorded an average HR deviation of just 1.8 bpm between deepfakes and their source videos, a value within the margin of error for many biometric systems.

Further complicating matters, some rPPG-based detection models have been found to mistakenly rely on background artifacts or lighting inconsistencies, rather than authentic blood flow analysis. This can lead to misclassifications, either failing to identify fakes or falsely flagging real content.

The research warns that deepfake detection must now evolve from analyzing global pulse presence to inspecting localized blood flow patterns. It proposes the use of spatially resolved rPPG signals that reflect anatomical consistency across different facial regions. Such localized maps could detect discrepancies in how blood flows across a fake face compared to a real one, offering a more nuanced and explainable detection strategy.

What comes next for deepfake detection?

The implications of these findings are profound. As deepfakes grow more realistic, detection systems must abandon overly simplistic assumptions and embrace multilayered, physiology-aware approaches. The study’s authors suggest that future detectors should integrate localized rPPG signal mapping, augmented with deep learning classifiers trained on anatomically consistent datasets.

They also stress the need for higher-quality datasets that include synchronized ECG or PPG ground truth for validation. Most public datasets, such as FF++ or Celeb-DF, are plagued by low resolution, compression artifacts, and noise, making them unsuitable for reliable rPPG extraction. To counter this, the researchers generated their own controlled dataset, capturing diverse subjects under consistent lighting and background conditions to ensure signal fidelity.

Furthermore, they propose using explainable AI tools to highlight exactly which facial regions or signal features lead to a detection verdict. Such transparency is essential for building trust in biometric verification systems and understanding the evolving dynamics of synthetic media.

The researchers warn of potential adversarial countermeasures. As detection tools grow more sophisticated, deepfake creators may attempt to explicitly model and embed localized blood flow patterns to evade scrutiny. This cat-and-mouse game between generation and detection will continue to escalate, but the study offers a clear roadmap for staying ahead.