Deepfake detectors rely on video cues beyond pulse
Pulse or Provenance? What rPPG-Based Deepfake Detectors Learn · arXiv:2611.13992
In plain words
Some deepfake detectors look for tiny color changes in a face that are linked to blood flow. This study asks whether those detectors actually use a heartbeat-like signal, especially for fake videos made by animating a still face.
The researchers compared color traces from matched real and fake videos. For talking-face fakes, a simple trace separated the videos better than a dedicated pulse extractor.
The signal also appeared in backgrounds and depended on frame timing, suggesting the tested detectors relied mainly on changes linked to video generation, not pulse.
The research
A detector can score highly while responding to how a video was generated or collected rather than whether it was manipulated. These findings make pulse-based scores harder to interpret as evidence of deepfakes, especially when videos come from different sources or evaluation splits reuse material.
The study tests what rPPG-style deepfake detectors learn and finds that temporal color artifacts, rather than pulse, drive much of their signal.
The researchers compared color traces from matched real and fake videos, testing different image regions, frame order and frequency bands. On talking-face fakes, the trace stayed strongly discriminative outside the face, but shuffling its frames brought performance near chance.
The strongest evidence favored temporal color artifacts over pulse: higher-frequency changes mattered more than the usual heart-rate band, and adding a synthetic pulse barely changed classifications.
Cross-source false alarms and a protocol test also showed that video origin and test-video reuse can shape scores.
The study has no ground-truth pulse measurements, so it cannot establish whether every fake lacks a pulse.
Its real test set has about 100 videos from nine identities, and the reenactment evidence covers two generators, with a frame-by-frame control for only one; video acquisition and processing also contribute to the cue.
What the evidence shows
- What do detectors learn?
Harraq and Aldwairi’s empirical study examines what remote photoplethysmography (rPPG)-style deepfake detectors learn from videos. These detectors are motivated by the idea that synthesized faces lack the blood-flow signal present in real skin.
The key test is whether their decisions actually depend on that signal, particularly for talking-face fakes made by animating a still image.
- A trace beats pulse extraction
On Celeb-DF++ talking-face test videos, the color trace separated real from fake more strongly than the dedicated pulse extractor. This gap suggests that strong detection scores do not by themselves show that a detector is measuring pulse.
- The cue extends beyond faces
For fully synthesized talking-face and face-reenactment videos, color traces from the background and full frame separated real from fake about as well as traces from the face.
Shuffling frame order removed that separation, indicating that the cue depends on changes over time but does not require facial skin.
- Higher frequencies drive decisions
On talking-face test videos, the higher-frequency color changes separated real from fake better than changes in the usual heart-rate range. Removing high-frequency content made most fake test videos appear real to the unchanged classifier, while adding a synthetic pulse had little effect. These tests point to video changes beyond pulse as the main cue in this setting.
- Generation and evaluation shape scores
Self-reenactments provide a check on where the signal can arise: animating from a single-frame makes a generated video distinguishable from its source, while a frame-by-frame reconstruction control is near chance. Yet detector outputs also track collection source: a detector trained on Celeb-DF++ calls 99.1% of held-out real FaceForensics++ videos fake, versus 1% of YouTube-real videos.
A separate protocol comparison shows test-video reuse inflates one weak detector’s AUC, so high scores alone do not show that a detector recognizes manipulation.
Deep
- Original paper authors
- Othmane Harraq, Tamer Aldwairi
- Submitted to arXiv
- License
- arXiv non-exclusive