Masks Look Old to a Pulse Clock: Age-Conditioned Pulse Morphology for Mask Detection and Vascular Age

A camera can see a heartbeat. Each pulse slightly changes how much light the facial skin absorbs, and face anti-spoofing systems already use this signal, called remote photoplethysmography or rPPG, to reject silicone and resin masks: no pulse, no person. These systems throw away the shape of each pulse once they confirm that a heartbeat frequency is present. Cardiovascular researchers, working with fingertip sensors, have shown that this shape encodes how stiff a person's arteries are, and therefore a "vascular age." We propose to keep the pulse shape and model it against age. The model learns what living pulse shapes look like at each age, using live faces only. The same model then does two jobs. It flags a presented face whose pulse shape is implausible for any living person of that apparent age, which is the signature of a mask worn over a living face. For faces that pass, it reports a vascular age gap: estimated age minus real age.

The idea rests on a physical observation. A thin mask worn by a living attacker does not remove the wearer's pulse. It attenuates the pulse and blurs it in time, because light scatters through an inert polymer layer before it returns to the camera. Blurring flattens the secondary, diastolic peak of the pulse. That flattening is the feature the fingertip age model of Nie et al. relies on to recognise old, stiff arteries. To a pulse-age model, a masked face should therefore look vascularly much older than its face looks, or fall off the range of living pulse shapes altogether. The detector needs only live training data, so it should hold up against mask materials it has never seen. A spoof-resistant age-gap measurement is also something the face-based biological age literature currently lacks.

Core Claim

Our central claim is that a normative model of facial pulse morphology conditioned on apparent age detects 3D masks better than pulse-presence features alone, especially on unseen mask materials. "Normative" here means fitted only to healthy, live faces across the adult lifespan, in the style of brain-age modelling. It also yields a camera-based vascular age that agrees with a fingertip-sensor clock on the same people.

Concretely, let $m$ be a fixed-length morphology descriptor of the averaged facial pulse (for example, the time-normalised mean beat plus its first derivative). Let $\hat a_f$ be the face's apparent age from an appearance model, and let $p_\theta(m \mid a)$ be a density over live pulse shapes at age $a$, learned on live faces only. We score a presentation by

$$ s(m, \hat a_f) = -\log p_\theta(m \mid \hat a_f) + \lambda \, \big| \hat a_p(m) - \hat a_f \big| , $$

where $\hat a_p(m)$ is the age the pulse-shape regressor assigns and $\lambda \ge 0$ is a weight tuned on validation data. High scores mean "not a plausible living face of this age." For accepted faces, $\Delta_p = \hat a_p - a_{\text{chron}}$ is the vascular age gap.

This differs from the closest work in specific ways. CFrPPG checks that a pulse frequency peak is consistent across face regions and stands out from background noise, using the peak-to-sidelobe ratio. Local rPPG perfusion assessment classifies masks from perfusion index and correlation maps with a support vector machine. Both are trained as discriminative live-versus-mask classifiers, and neither models pulse shape or age. AI-PPG age models pulse shape and age but uses contact sensors and never considers attacks. FaceAge estimates biological age from still photographs and has no defence against a physical mask. Our proposal is the first to put age-conditioned pulse shape inside a mask detector.

Research Program

The work has three stages, each with a clear failure condition.

Stage 1: transfer a pulse-age clock from fingertip to face. Camera pulses are far noisier than contact pulses, and rPPG-Toolbox mainly benchmarks heart-rate accuracy rather than waveform shape. So the first task is to learn a mapping from facial rPPG beats to contact-sensor beats, using datasets in that toolbox that record both (UBFC-rPPG, PURE). We extract beats with POS, a classical method that projects colour changes onto a plane orthogonal to skin tone, and with PhysNet, a deep pulse-extraction network. We average beats over windows of about ten seconds, which matches the contact recordings Nie et al. used.

Those public datasets are dominated by young adults, so age coverage requires a new collection. We plan roughly 200 adults aged 20 to 80, recorded with synchronised RGB video and a finger sensor, at rest, under two lighting conditions. We train a Net1D-style regressor with Nie et al.'s distribution-aware loss. That loss counters the tendency of age models to pull every prediction toward the middle of the age range. We train first on contact pulses and then fine-tune on mapped camera pulses. The density $p_\theta$ is a conditional normalising flow, a generative model that gives exact likelihoods, fitted on live beats only.

Stage 2: same subjects, with masks. A subset of about 60 participants wears thin transparent masks and thicker resin or plaster masks of the kinds used in HiFiMask. This produces matched recordings of the same wearer with and without a mask. We can then measure directly whether masks shift $\hat a_p$ upward and lower the live-pulse likelihood, as the blurring argument predicts. Public mask sets (HKBU-MARsV2, 3DMAD, HiFiMask, and FaceMask if obtainable) serve as external tests. Most of them lack subject ages, so on these sets we use $\hat a_f$ in place of real age. That is the realistic deployment setting anyway.

Stage 3: an exploratory epigenetic link. Here the epigenetic direction plays a modest, explicitly secondary role. The regenerative-barrier literature describes how adult tissue loses its capacity to remodel: cardiomyocyte cell-cycle arrest enforced by non-coding RNA networks, and chromatin compaction that blocks stem-cell repair until opened by NuRD remodelling. FaceAge's association with the senescence-linked gene CDK6, and methylation age acceleration in chemotherapy patients, suggest that visible ageing has molecular correlates. In consenting Stage 1 participants, we would draw blood for a DNA-methylation age estimate. We would then test whether the camera vascular gap $\Delta_p$ correlates with methylation age acceleration beyond what FaceAge's appearance gap explains. This tests whether the optical clock reflects something biological rather than sensor noise. It does not test barrier mechanisms directly, and the main claims do not depend on it.

Baselines

The naive reference is a single-region green-channel peak-frequency liveness check. Liu et al. report that such naive rPPG methods reach error rates as high as 31% on RAD. The texture reference is MS-LBP. In the cross-dataset protocol it reaches 39.8% equal error rate (EER, the error rate at the threshold where false accepts equal false rejects) when trained on 3DMAD and tested on HKBU-MARsV2. These two numbers bound the low end.

The strongest pulse-based competitor is CFrPPG. It addresses the same question with the same signal, and it reports 4.04% EER intra-dataset on HKBU-MARsV2, 5.08% EER cross-dataset from 3DMAD to HKBU-MARsV2, and 6.17% EER on RAD. We will re-run it through the same face tracking, the same window lengths, and the same splits. Our method must at minimum match these figures. Beating CFrPPG means a lower EER on our same-subject masked collection and on HiFiMask leave-one-material-out, with the improvement concentrated on thin transparent masks, where pulse-presence features should struggle.

The local perfusion competitor is Kossack et al.'s perfusion index and correlation features with an SVM, which reports 96.8% accuracy on FaceMask. We rerun it on our data with an equal tuning budget.

The strongest appearance competitor is the HiFiMask challenge winner. VisionLabs' part-based multi-branch network reports ACER (the average of the two classification error rates) of 3.053% and AUC of 0.995 on HiFiMask, and the runner-up's central-difference convolution network reports ACER 3.155%. We will train a central-difference convolution backbone on the same HiFiMask split with the same compute budget. We expect it to remain hard to beat when mask materials are seen in training. The contest is on held-out materials.

For age estimation, the reference point is the contact-PPG clock of Nie et al., with reported mean absolute error of 7.5 to 11.8 years. A camera clock within roughly two years of the contact clock's error on the same subjects counts as success. Predicting the cohort mean age is the do-nothing floor.

Two ablations isolate the key ingredient. The first removes age conditioning, using an unconditional live-pulse density. The second removes shape and keeps only the peak-to-sidelobe ratio and perfusion features inside the same one-class model.

Why not just…?

Why not just fine-tune CFrPPG or another pulse-presence detector on more masks? For thick masks this is a real threat, because there may be no transmitted pulse to analyse and presence alone may suffice. Our argument concerns thin and transparent masks over living wearers, where a pulse does leak through. There, a discriminative classifier has to have seen that particular material during training. The leave-one-material-out experiment and the shape-removed ablation decide this question. If peak-to-sidelobe features match our full model on held-out thin masks, the shape claim fails and we will report that.

Why not just train an end-to-end video network and let it learn pulse shape implicitly? Strong appearance models already win HiFiMask, and given seen materials they may match us. They are discriminative, however: they learn what training masks look like, not what living skin does. We expect a gap only on unseen materials and capture conditions. If the central-difference convolution baseline matches us there too, the remaining contribution is the calibrated vascular age output alone, and we would present the work as such.

Why not just run FaceAge plus a separate liveness check? That pipeline yields an appearance age, not a vascular one. FaceAge is also exactly what an "old-looking" mask is designed to fool, while the pulse is not. Stage 3 measures what the vascular gap adds beyond the appearance gap. The age-mismatch term in $s$ tests whether combining the two detects more attacks than either does alone.

Proposed Experiments and Decisive Checks

Three results decide the project. First, in matched same-subject recordings, masks must shift pulse age upward or lower live-pulse likelihood. If masked and unmasked beats are indistinguishable after mapping, the mechanism is absent and we stop. Second, on HiFiMask leave-one-material-out and on HKBU-MARsV2 cross-dataset, the full model must beat CFrPPG's 5.08% cross-dataset EER under our reimplementation and beat the shape-removed ablation. Third, the camera pulse age must correlate with the contact Net1D age of the same subjects, with error within about two years of the contact clock's. Without that, the "vascular age" output is noise.

We also report EER and ACER under dim and backlit HiFiMask conditions, and accuracy as a function of window length from 3 to 10 seconds. The window-length analysis is the honest cost of this approach: the method needs several seconds of video and is not suited to single-shot authentication. The methylation correlation is reported as exploratory, with its sample size stated.