Pulse-based deepfake detectors mostly spot flicker and video source, not heartbeats

A preprint finds that deepfake detectors built to read a heartbeat from facial skin colour mainly pick up frame-to-frame colour flicker. They also learn which collection a video came from. Common benchmark splits can hide this.

6 min readIIBenchmarkedPreprint, not peer reviewed
Save
A brass stethoscope on warm paper, with its chestpiece resting on a deep blue luggage tag beside film canisters.AI illustration
Illustration: AI generated

What the study shows

Remote photoplethysmography (rPPG) reads tiny colour changes in facial skin caused by blood flow. Several deepfake detectors rely on it. Their premise is simple: a synthetic face has no blood flow, so it should have no pulse. The authors test what these detectors actually pick up. They use Celeb-DF++, a public benchmark with 22 forgery methods, and FaceForensics++ (FF++), an older benchmark. Each real and fake video gives exactly one window: its first 160 frames.

The authors report that on talking-face fakes, where a still photo is animated from audio, a crude colour signal from the face almost perfectly separates real from fake. Its AUC is 0.998. AUC is a score where 1.0 is perfect and 0.5 is a coin toss. A dedicated pulse-extraction model scores only about 0.87. If the cue were a heartbeat, the better pulse estimator should win.

The face turns out not to matter. On talking-face and face-reenactment fakes, a patch of background, or the whole frame with no face detection, works as well as the face. The useful information sits above the normal heart-rate range. Shuffling the order of frames destroys it. The cue is how colour changes from one frame to the next.

Targeted edits confirm this. Filtering out fast colour changes made 97% of fakes look real to the detectors. Adding fake-like fast changes to real videos made almost all of them look fake. Adding an artificial pulse to fakes made almost no difference. The cue also survived matching frame rate, codec and resolution between real and fake videos.

To find the source, the authors regenerated real videos from themselves with two face-animation tools. Animating a face from a single frame added the same flicker, and the fakes became separable from their own originals. Rebuilding each frame from itself did not add it. This second control was run with one tool only.

The detectors also learned video source. Trained on Celeb-DF++, colour-based detectors called over 99% of real FF++ videos fake. A general-purpose detector built on CLIP, an image model trained on pictures paired with captions, called most of them fake too. Adding pulse features to CLIP did not improve it.

Finally, the official Celeb-DF and Celeb-DF++ test lists share people and source videos with the training data. In one custom protocol, reusing test videos in training raised a weak detector's score by about 0.30 AUC. A strong detector gained only about 0.05.

Our reading

In our reading, this is a careful audit, not a new detector. Its main value is the set of controls. The authors test a background patch, shuffled frames, separate frequency bands and identical encoding. They also run an independent reimplementation with no shared code. Together these make the conclusion hard to dismiss. A detector justified by biology is largely detecting a rendering artefact and a dataset fingerprint.

The practical risk is false alarms. Suppose a newsroom uses such a detector to check a genuine clip from a phone or a platform it never saw in training. On this evidence, the tool could label that clip fake with high confidence. The FF++ result is exactly this case: real videos, scored almost uniformly as fake.

It is also notable that part of the critique targets the authors' own earlier pulse-based work. The paper is a preprint and has not been peer reviewed. On our four-level readiness scale it sits at level II, benchmarked: results are on public benchmarks. Code and evaluation splits are promised but not yet released, so it does not reach level III.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

II
Benchmarked
LevelStatusEvidence
I  Lab onlyMetThe work is a research analysis with controlled lab experiments on public datasets.
II  BenchmarkedMetResults are reported on public benchmarks (Celeb-DF++ with 22 forgery methods, and FF++), with AUC, confidence intervals, 5 seeds and an independent reimplementation. This is an analysis of detectors rather than a new detector.
III  Open sourceNot yetThe paper says code and evaluation splits 'will be made publicly available'. No link is given, so nothing is released yet.
IV  In productionNot yetNo documented use by a named organisation in production. The work is a diagnostic study, not a deployed system.

What would move it up
Release the promised code and evaluation splits to reach level 3. Peer review and replication by other groups on newer generators and other datasets would strengthen the findings. Level 4 does not clearly apply, because this is an audit method rather than a product.

Checked automatically every week, last on 16 November.

The study

Othmane Harraq, Tamer Aldwairi. Pulse or Provenance? What rPPG-Based Deepfake Detectors Learn. arXiv, 10 November 2026. doi:10.48550/arxiv.2611.13992

Preprint, not peer reviewed · Licence arXiv (no reuse) · Temple University, Philadelphia

Notes

  1. Harraq, O. and Aldwairi, T. Pulse or Provenance? What rPPG-Based Deepfake Detectors Learn. arXiv:2611.13992, 10 November 2026. Preprint, not peer reviewed. Source
  2. Code and evaluation splits: promised by the authors but no link given at time of writing.
  3. Key figures on talking-face fakes: face CHROM AUC 0.998, background 0.998, full frame 0.999, RhythmFormer 0.869. 4 to 12 Hz band alone 0.992 against 0.943 for 0.7 to 4 Hz. Shuffled frames 0.521.
  4. Interventions: low-pass at 8 Hz flips 97% of fakes to real. Fake-level 12 to 12.5 Hz noise flips 99.8% of reals to fake. A synthetic pulse leaves 2.5% of fakes called real, against a 2.0% baseline.
  5. Self-reenactment: LivePortrait up to 0.931 AUC from the first frame. Frame-by-frame rebuild 0.582.
  6. Source encoding: face CHROM trained on talking-face Celeb-DF++ calls 99.1% of real FF++ videos fake. CLIP 86.0% (talking-face training) and 87.8% (all 22 methods). Strict protocol: CLIP 0.957, CLIP with rPPG and DCT fusion 0.938.
  7. Leakage: rPPG+DCT detector 0.929 under a leave-one-method-out protocol that reuses test reals, against 0.634 with identities held out. CLIP 0.994 against 0.947.
  8. Readiness: level II, benchmarked. Public benchmark results reported. No released code (level III) or documented production use (level IV).

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Safety & evaluation