Remote 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 0.998, exceeding a dedicated rPPG extractor (0.87). Controlled experiments indicate that this signal is not physiological in origin. On fully synthesized forgeries, it is equally discriminative when extracted from background regions or from full frames without face detection, and it is concentrated above the 0.7 to 4 Hz heart-rate band. It is largely removed by shuffling frame order but persists after normalizing resolution, frame rate and encoding, and an independent reimplementation reproduces these findings. Removing high-frequency content makes most synthesized videos appear real, adding it makes real videos appear fake, and injecting a synthetic pulse has little effect. In controlled self-reenactment experiments with two generators, animating a face from a single frame introduces the same high-frequency signal, whereas reconstructing each frame from its own source does not. Face-based chrominance detectors trained on Celeb-DF++ classify over 99% of real FF++ videos as fake, indicating that they largely encode video source rather than manipulation, and a CLIP-based detector exhibits the same behavior. A frozen CLIP backbone matches or exceeds fusions of CLIP with rPPG and frequency features. Finally, we show that the official Celeb-DF and Celeb-DF++ splits share identities and source-video content between training and test, and that custom protocols reusing test videos and identities in training inflate weaker detectors by up to 0.30 AUC.
Figures & tables
Fig. 1: rPPG-style detectors do not need a face or a heartbeat. Top: real vs. fake AUC by region and signal. Bottom: face CHROM with frames shuffled or restricted to one frequency band. On talking-face (TF) and face-reenactment (FR) forgeries, the background and the full frame separate real from fake as well as the face, the 4–12 Hz band beats the 0.7–4 Hz heart-rate band, and shuffling frames removes the signal. Matched sampling, test identities; dashed line is chance.
Set
Signal
Face
Background
Full frame
Shuffled
0.7–4 Hz
4–12 Hz
TF (384/98)
CHROM
0.998
0.998
0.999
0.521
0.943
0.992
POS
0.998
0.993
0.998
0.593
0.897
0.970
Green
0.931
0.960
0.959
0.604
0.802
0.911
RhythmFormer
0.869
–
–
0.679
0.860
0.800
FR (315/99)
CHROM
1.000
0.998
1.000
0.505
0.809
0.996
POS
1.000
0.995
1.000
0.493
0.740
0.994
TABLE I: Real vs. fake AUC under matched sampling (one window per video, test identities, mean of 5 seeds). Shuffled and band columns use the face signal. Bands are applied before per-window normalization. TF uses the fps25 variant. Counts are fake/real test videos. RhythmFormer operates only on detected face crops, so it has no background or full-frame entries (–).
Fig. 2: Averaged power spectra of real (black) and fake (blue) CHROM traces under matched sampling. Rows: face, background, full frame. Columns: forgery category. The shaded band is 0.7–4 Hz. Real and fake spectra separate above 4 Hz for TF and FR, only slightly for FS, and not at all for FF++.
Variant
Plain
0.7–4 Hz
4–12 Hz
Native files
0.998
N/A
N/A
fps25 (frame rate, codec matched)
0.998
0.943
0.992
res_norm (face crop 256 2 , 30 fps)
0.998
0.859
0.993
res25 (crop, resolution, fps, codec)
0.986
0.906
0.978
Reimplementation, face
0.989
0.802
0.983
Reimplementation, background
0.995
0.698
0.952
TABLE II: TF face CHROM under progressively stronger normalization (matched sampling), and an independent reimplementation (logistic regression on the trace spectrum, no shared code). Band columns are N/A for native files, whose frame rates differ between classes (25 vs. 30 fps), and for shuffled traces, which have no temporal structure to filter.
Intervention
AUC
Flipped
None
0.998
2.0% of each class
Fakes low-passed at 8 Hz
0.589
97% of fakes
Fakes low-passed at 4 Hz
0.877
85% of fakes
Fakes, 12–12.5 Hz removed
0.825
75% of fakes
Fakes, 4–12 Hz removed
0.999
1.6% of fakes
Reals + 4–12 Hz noise (fake level)
0.998
1% of reals
TABLE III: Interventions on test traces with unmodified TF face-CHROM models (fps25, matched sampling). Flipped: fraction of the modified class on the wrong side of the unmodified models’ EER threshold. Low-pass filters also remove content up to the 12.5 Hz Nyquist limit.
Fig. 3: Real videos vs. their own self-reenactments. Top: face CHROM AUC for separating each reenactment from its source video, plain, shuffled, and per frequency band. Bottom: share of face CHROM power in 4–12 Hz for each real video and its reenactments. (a) animated from the first frame and pasted back; (b) generated face crop; (a2) each frame rebuilt from itself. LP: LivePortrait.
Method
nfake
Face
Bg
Frame
RF
FS
BlendFace
55
0.945
0.772
0.911
0.534
Celeb-DF
183
0.604
0.509
0.629
0.473
GHOST
55
0.973
0.756
0.923
0.502
HifiFace
55
0.953
0.769
0.907
0.539
InSwapper
55
0.954
0.749
0.905
0.482
MobileFaceSwap
55
0.721
0.507
0.695
0.490
TABLE IV: Per-method real vs. fake AUC under matched sampling (mean of 5 seeds; test identities; about 98–139 real test videos per set). Face, Bg and Frame are CHROM traces from the face, background and full frame; RF is RhythmFormer. † EchoMimic outputs 24 fps and is scored on native files. Celeb-DF denotes the original Celeb-DF-v2 forgeries.
Training
Model
YouTube
FF++
Celeb vs FF++
TF
Face CHROM
1%
99.1%
0.998
TF
Background CHROM
3%
77.8%
0.959
TF
CLIP
38%
86.0%
0.948
All 22
Face CHROM
18%
99.7%
0.999
All 22
CLIP
47%
87.8%
0.945
TABLE V: Real videos from other sources, scored by detectors trained on Celeb-DF++ (matched sampling). Called fake: fraction above each model’s own EER threshold. Last column: AUC separating held-out Celeb-real from FF++ real using the fake score.
Fig. 4: The detectors encode video source. Top: fraction of each real collection called fake by detectors trained on the TF subset or on all 22 Celeb-DF++ methods, at each detector’s own EER threshold. Bottom: classifiers trained to separate two collections of real videos, with a random-label control.
Model
FS
FR
TF
All
Face CHROM
0.877
1.000
0.992
0.951
CLIP (frozen)
0.971
0.956
0.943
0.957
Xception (frozen)
0.673
0.841
0.806
0.765
RhythmFormer
0.499
0.634
0.727
0.614
Strict
GF-eval
CLIP only
0.957 [0.946, 0.968]
0.885 [0.877, 0.891]
TABLE VI: Top: detectors on the same strict identity-disjoint protocol, matched sampling, 95 real and 1,301 fake test videos shared by all four detectors, mean of 5 seeds. Bottom: CLIP vs. fusion on the matched feature cache, with 95% CIs; strict uses a fixed 30 epochs, as in Table VIII . GF-eval follows the benchmark protocol, in which about 90% of test reals are also used in fine-tuning; with only the official test reals, the three rows give 0.844, 0.830, and 0.513.
Celeb-DF++
Celeb-DF-v2
FF++
Test fakes sharing an identity
100%
100%
– ∗
FS/FR test fakes sharing a pair
100%
100%
0%
FS/FR test fakes on a training real
78–80%
69%
0%
TF test fakes from a training real
85%
N/A †
N/A †
Test reals used for training fakes
92–100%
90%
0%
TABLE VII: Overlap between training and test in the official splits (test side). ∗ FF++ has no identity labels; its split is defined by source videos. † No TF subset.
Protocol
rPPG+DCT
CLIP
Fusion
Random video split
0.680
0.976
0.974
LOMO, loose
0.929
0.994
0.996
LOMO, clean reals
0.600
0.971
0.967
LOMO, strict
0.634
0.947
0.929
Identity-disjoint, strict
0.680
0.957
0.938
TABLE VIII: Leave-one-method-out (LOMO) and identity-disjoint protocols on Celeb-DF++, matched sampling. Loose: identities and real videos shared. Clean: official test reals held out. Strict: identities held out. LOMO values are the mean AUC over the 21 held-out methods.
Recent deepfake detection studies increasingly suggest remote photoplethysmography (rPPG) signals as an authenticity cue. However, existing benchmarks lack physiological ground truth, and current detectors underexplore the cross-level relationship between facial features and physiological dynamics, often relying on late fusion or rPPG features alone. In this paper, we construct high-fidelity deepfake manipulations on established real rPPG datasets (COHFACE and UBFC-rPPG) to investigate how forgeries disrupt natural physiological signals and facial behavior at the same time. Building on this analysis, we propose a bidirectional co-attention fusion detector that jointly models rPPG and facial behavior tokens. This mechanism explicitly captures the cross-level dependencies between pulse dynamics and facial motion to learn a robust, joint authenticity representation. Extensive experiments using a subject-disjoint 5-fold evaluation demonstrate the superiority of our approach. Achieving a 92.80% AUC on constructed datasets using face swapping and 96.78% AUC on motion transfer, our model outperforms both the rPPG-only single modality baseline and the best feature-level fusion methods. Furthermore, transfer-learning result of the fusion detector on Celeb-DF-v2 while keeping both feature extractors fixed achieves 91.20% accuracy and 86.08% AUC, which suggests applicability under target-domain adaptation.
Chenxi Yang, Yassine Ouzar, Larbi Boubchir
LIASD Laboratory Shanghai Jiaotong University Shanghai, China · LIASD Laboratory University of Paris 8 Saint-Denis, France
Talking-face (TF) deepfake generation synthesizes photorealistic facial video from a static source image and an audio signal, producing forgeries that current image-based detectors consistently fail to identify. Unlike face-swap manipulation, TF synthesis has no underlying real video from which to inherit physiological characteristics, motivating remote photoplethysmography (rPPG) as a detection modality for this forgery category. We propose PulseGuard, a detection framework that extracts per-window rPPG waveforms via RhythmFormer and trains a suite of lightweight classifiers to distinguish real from synthesized traces. Evaluated on the TF subset of Celeb-DF++ under a strict subject-independent protocol, where test identities are completely separated from training identities, our 1D ResNet achieves a window-level AUC of 0.826 at EER 25.1% (0.842 pooled to video level, the granularity at which DeepFakesON-Phys is scored) while operating exclusively on the extracted temporal signal, and retains AUC 0.754 on generators withheld entirely from training. We document a controlled reproduction study of DeepFakesON-Phys, the representative prior rPPG detector, demonstrating degradation from AUC 0.999 on legacy face-swap data to 0.576 on the TF subset of Celeb-DF++. We further show that detection difficulty is strongly method-dependent: AUC ranges from 0.937 (Real3DPortrait) to 0.617 (IP-LAP) across the seven TF generators, with the ordering broadly preserved under per-method training. This spread is stable across training regimes rather than reflecting evaluation noise, and constitutes the primary empirical contribution of the work. We report these as empirical baselines and do not establish that the discriminative content is cardiac in origin.
Talking-face (TF) deepfakes are detected unevenly by rPPG-based methods across generators. We study two lightweight visual-only cues, rPPG-derived waveforms extracted by RhythmFormer and lip-region discrete cosine transform (DCT) coefficients, on the seven TF methods of Celeb-DF++ under a subject-independent protocol. In-domain, lip-region DCT matches or exceeds the rPPG-derived 1D ResNet on every method except SadTalker, and Concat fusion reaches AUC 0.891 against 0.824 and 0.827 for the unimodal baselines. Under leave-one-generator-out evaluation the cues split: each transfers clearly better to three held-out methods, and IP-LAP is near chance for both. Concat averages 0.798 but falls below rPPG alone where DCT transfers poorly, so static fusion only partly exploits this complementarity. Lip-region DCT outperforms full-face DCT on six of seven methods. We treat the rPPG-derived signal as an empirical cue and do not claim it is cardiac in origin.