rPPG deepfake detectors track high-frequency artifacts, not pulse
Original title: Pulse or Provenance? What rPPG-Based Deepfake Detectors Learn
Chrominance traces hit AUC 0.998 on talking-face forgeries, and Celeb-DF splits leak identities into the test set.
A study tests what remote-photoplethysmography deepfake detectors learn on Celeb-DF++ (22 forgery methods) and FaceForensics++, using one identically sampled window per real and fake video. On talking-face forgeries, simple chrominance traces reach AUC 0.998 versus 0.87 for a dedicated rPPG extractor, and controlled tests indicate the cue is not physiological. The signal sits above the 0.7–4 Hz heart-rate band, survives resolution, frame-rate and encoding normalization, and is reproduced by an independent reimplementation. Face chrominance models trained on Celeb-DF++ label over 99% of real FF++ videos as fake, and the authors say official Celeb-DF and Celeb-DF++ splits share identities and source video between train and test.
- Shuffling frame order largely removes the signal; a synthetic pulse has little effect.
- The cue remains discriminative from background regions and from frames without face detection.
- Removing high-frequency content makes most synthetic videos look real.
- Custom protocols that reuse test videos inflate weaker detectors by up to 0.30 AUC.
- A frozen CLIP backbone matches or exceeds fusions with rPPG and frequency features.
- Products
- Celeb-DF++FaceForensics++
- AI models
- CLIP
From the original article
1 min read at arxiv.orgRemote photoplethysmography (rPPG) has been proposed as a physiological cue for deepfake detection, based on the premise that synthesized faces lack a genuine cardiac signal. We examine what rPPG-style detectors learn on Celeb-DF++ (22 forgery methods) and FaceForensics++ (FF++), using one identically sampled window per real and fake video. On talking-face forgeries, simple chrominance traces reach an AUC of 0.
By Othmane Harraq · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.