A Mask with a Manufactured Pulse: Photo-Switchable Host-Guest Films Against Liveness Detection

Face-verification systems increasingly rely on a simple physiological fact to tell a person from a photograph or a replica: your heartbeat. Every cardiac cycle pushes blood through capillaries just under the skin, producing a faint, rhythmic color change — remote photoplethysmography, or rPPG, reads this change from ordinary camera video and rejects anything that does not pulse. This defense has held so far because every known 3D mask material — resin, plaster, silicone, wax — is inert: it can copy a face's shape and color, but it cannot copy its blood flow. We propose to build and evaluate the first presentation attack that closes this gap, not with electronics, but with a passive molecular film.

The idea in two sentences: coat a 3D facial mask with a photo-responsive host-guest polymer — a material whose molecular crosslinks open and close reversibly under ultraviolet and visible light, changing the film's optical density and heat output — and drive it with a small, camera-invisible ultraviolet light source modulated at a heartbeat rhythm. The mask then emits a locally generated, spatially patterned chrominance and thermal oscillation that mimics a cardiac signal, and the research question becomes quantitative: how close can a physical artifact get to statistical indistinguishability from a live face in the channels a defender can observe?

Core Claim

Current 3D-mask defenses split into two families that fail for different reasons. Material classifiers, which separate skin from polymer by reflectance, collapse when confronted with unfamiliar materials: the intrinsic-image-analysis method of Li et al., which separates each face frame into a shading component and a material-reflectance component and classifies the latter, achieves a 0.0% equal error rate on the 3DMAD database of resin masks but degrades to roughly 50% average classification error when tested on a different mask corpus. Liveness detectors bypass material entirely by looking for the pulse: the rPPG correspondence-feature method of Liu et al., which learns a spatial template of the pulse spectrum across facial subregions and subtracts background illumination noise as hard negatives, reaches 4.04% EER on the HKBU-MARsV2 benchmark of commercial hyper-realistic masks.

Our claim is that this second family, the last robust line of defense, is attackable with passive switchable matter, and that the attack can be made quantitative by borrowing the defendability game formalism of Christiano et al.. They frame backdoor defense as a discrimination game — a defender must tell a clean function-with-input from a backdoored one — and prove that if the attacker can make the two computationally indistinguishable (via cryptographic indistinguishability obfuscation, iO — a compiler that transforms two functionally identical programs into forms no efficient analyzer can tell apart), then defense is impossible. We transplant this lens from circuits to physical observables: the spoof succeeds to the degree that the joint distribution of the mask's visible chrominance and long-wave infrared emission over time becomes statistically indistinguishable from a live face's. We cannot and do not claim cryptographic iO for matter; we claim an empirical measurement of how far a real material can push toward that limit, and of which observable cues survive — which is exactly the information the next generation of defenses needs.

Role of Each Prior Line

Three strands supply the mechanism. The azobenzene host-guest chemistry provides fast, reversible optical switching: in the interpenetrating-network hydrogel of Li et al., an azobenzene guest snaps between a straight form that sits inside a cyclodextrin host cavity (crosslink engaged, film dense) and a bent form that is expelled (crosslink released), shifting the swelling ratio from 72% to about 97% and restoring shape in about one minute under UV; in the azobenzene inverse-opal film of Liu et al., the same isomerization raises the effective refractive index from 1.196 to 1.329 and flips the film's structural color within two seconds of illumination, reversibly for over seventy cycles. The ternary nanoporous network of Shen et al. shows the same trap-and-release logic at the single-cavity scale, confirming the mechanism is not tied to one scaffold. Color-driven thermal modulation supplies the infrared design tool: CDUPatch demonstrated that surface color alone sets the sunlit equilibrium temperature through the balance

$$\alpha_{\text{solar}} I_{\text{solar}} = h_c (T_{\text{surf}} - T_{\text{amb}}) + \epsilon \sigma (T_{\text{surf}}^4 - T_{\text{sky}}^4),$$

where $\alpha_{\text{solar}}$ is solar absorptivity, $I_{\text{solar}}$ incident irradiance, $h_c$ the convective coefficient, $\epsilon$ emissivity, and $\sigma$ the Stefan-Boltzmann constant — and trained a small MLP that maps any RGB pixel to its steady-state thermal signature, achieving a 70.4% attack success rate against YOLOv3 on DroneVehicle where the prior best managed 46.9%. We reuse this adapter idea, but time-resolved: our film's absorption oscillates rather than sits at equilibrium, so the adapter must map a driving waveform to a thermal waveform, and it must work under indoor illumination rather than sunlight (CDUPatch's stated failure mode).

The cryptographic strand plays two roles. The undetectable-backdoor pipeline of the Insidious Pipeline paper motivates the framing — it proved that obfuscation makes digital backdoors white-box undetectable and non-replicable via a signature check, but remains purely theoretical with prohibitive overhead and no physical triggers — while Christiano et al. supply the evaluation formalism and its limit theorem. One of their positive results also disciplines our design: their framework assumes the trigger is sampled randomly, so a worst-case engineered trigger is precisely the open case their proofs do not cover — our mask is exactly such a trigger, chosen adversarially.

Research Program

The first workstream builds the material. We will fabricate a conformal azobenzene host-guest film — starting from the acrylamide-azobenzene / polycyclodextrin chemistry of the hydrogel work, crosslinked as a thin skin rather than a bulk gel — over a 3D-printed facial replica, with regionally varied host density so that different facial subregions (cheeks, forehead, nose) can carry different modulation amplitudes, matching the spatially non-uniform pulse template that CFrPPG-based detectors expect. The key engineering risk is bandwidth and spectral purity: rPPG detectors look for a quasi-periodic signal near 1 Hz, and the inverse-opal film demonstrates 2-second switching, which is marginal. We will therefore characterize the achievable modulation depth and waveform fidelity in both the green-channel chrominance (the rPPG channel, where hemoglobin absorption lives) and the long-wave infrared band, measuring whether isomerization-driven changes in absorption and non-radiative heating produce a temperature oscillation within a thermal camera's noise-equivalent temperature difference. The UV driver is deliberately chosen to be invisible in RGB video, so the actuation light itself never appears as a global illumination flicker that a noise-subtracting detector could cancel.

The second workstream designs the driving waveform. Rather than guess a synthetic pulse, we optimize it: record paired RGB and LWIR video of live subjects, extract their per-region chrominance and thermal pulse spectra, and search (with expectation-over-transformation style optimization over waveform phase, amplitude, and inter-region delays, in the spirit of CDUPatch's robustness training) for the driving pattern that minimizes the distance between the mask's observable feature distribution and the live one, as scored by the CFrPPG feature extractor itself. This is the physical analogue of compiling a backdoored circuit to be indistinguishable from an honest one.

The third workstream is the evaluation and the theory. We instantiate the defendability game over observable video streams: the defender sees either a live-subject clip or a mask clip and must classify; we sweep the attacker's fidelity (modulation depth, spectral match, spatial template accuracy) and trace the defender's error from its strong baseline toward the 50% chance level. The residual gap — whatever the defender still exploits at high attacker fidelity, such as thermal diffusion phase lag or harmonics outside the cardiac band — is the constructive output of the project: a ranked list of cues that switchable-matter attacks cannot easily erase.

Baselines

The naive reference point is a conventional inert 3D mask with no dynamic element: the same physical replica, statically finished with tinted silicone paint to match skin color. Against liveness detection this should fail almost completely, because the rPPG detector of Liu et al. achieves 4.04% EER on HKBU-MARsV2 against precisely such masks, and 5.08% EER even when trained on a different database (3DMAD) and tested cross-corpus, where multi-scale local binary pattern baselines degrade to 39.8% EER. Running this arm establishes the floor and confirms that material appearance alone is not what our attack exploits.

The strongest existing competitor on this question is CFrPPG itself, the rPPG correspondence-feature detector: it learns a spatial pulse-spectrum template by ridge regression over facial subregions, uses background regions as hard negatives to remove ambient flicker, and scores each clip by the peak-to-sidelobe ratio against the template. It is the strongest because it is the one published method designed to be robust to exactly what static material engineering provides — realistic appearance — and it carries the best reported numbers on the hardest mask corpus (4.04% EER on HKBU-MARsV2, 6.17% EER on the Replay Attack Dataset under severe motion and low light, where competing rPPG methods reach up to 31% EER). We will retrain it from its published protocol on our data splits, give it the same video lengths and preprocessing it was designed for, and — to make the comparison adversarially fair — also give a defender-side collaborator white-box access to our mask clips so they can fine-tune or adapt the detector against us. Our attack beats this baseline only if the detector's EER on mask clips rises substantially above its published operating point, with the decisive threshold being the approach of 50% EER (chance) at full attacker fidelity; any rise short of that still quantifies partial breakage.

Two further baselines bracket the problem. On the material side, the contrastive context-aware learning framework of Liu et al. — which pairs live and masked faces of the same subject under matched illumination to isolate material differences, reducing a ResNet50 baseline's ACER on CASIA-SURF HiFiMask from 4.7% to 2.4% — represents the best material classifier and will be run on our mask both with switching on and off. On the thermal side, a static color patch designed with the CDUPatch pipeline serves as the thermal-modulation baseline; it is the strongest prior method for controlling infrared appearance without electronics, but its dependence on solar illumination means no reported result found on Lacuna for its indoor or nighttime effectiveness, which is itself part of what we measure.

The key ablation isolates the cardiac matching as the active ingredient: the identical switchable mask, driven by a waveform with matched total modulation energy but non-cardiac content (band-limited noise outside 0.7–1.5 Hz, or a cardiac-rate signal with the wrong spatial template across subregions). If the ablated mask fools CFrPPG nearly as well as the matched one, then the detector is broken by any temporal activity and our specific pulse-emulation claim is wrong; the proposal's central hypothesis requires a large gap between the matched and ablated conditions.

Why not just…?

Why not just make a better static mask — pre-tinted resin, layered silicone paint, wax? This is the incumbent approach, and the evidence says it caps out against liveness detection regardless of craftsmanship. The intrinsic-image method of Li et al. catches even well-matched static albedo at 0.0% EER on its training corpus, and more decisively, the rPPG detector of Liu et al. does not look at material at all — an inert mask has no pulse to fake, so no static improvement moves that needle. Static work also shows the fragility we want to avoid: cross-dataset material transfer collapses to roughly 50% ACER. The only un-attacked axis is the temporal physiological one, which is why the dynamic mechanism is the point rather than a complication.

Why not just put a flexible screen or LED strip under a translucent mask? This is the most serious alternative, and we treat it as a real threat rather than argue it away. Active electronics demonstrably modulate thermal appearance — the thermoelectric display of Li et al. reaches a 96.1% physical attack success rate against YOLOv5 at 2 meters, far outperforming passive baselines — and a thin OLED could in principle pulse visible chrominance at 1 Hz. We will therefore build a flexible-LED mask arm into the physical comparison and report it symmetrically. Our honest expectation, grounded in the literature, is that the material wins on detectability rather than raw spoof rate: powered hardware is rigid, needs a battery, emits its own heat signature that a thermal sensor sees, and presents electronics detectable by the multi-sensor, seven-device protocols that the HiFiMask challenge of Liu et al. showed are already standard, whereas a passive film is conformal, unpowered, and switches only when the attacker's covert UV source is present. If the LED arm matches the film's spoof success and survives the multi-sensor checks, the material claim weakens to a cost and detectability argument, and we will say so.

Why not just shine a pulsing visible light or projector at the face? Global illumination flicker is the cheapest trick, and it is specifically the one the strongest baseline was built to cancel: CFrPPG extracts signals from non-face background regions as hard negatives precisely to subtract ambient flicker and camera-motion noise, and the literature already notes that periodic ambient lighting can induce pseudo-pulse signals — which is why the detector's background-reference design exists. A projector bathes the background in the same flicker it bathes the face in, handing the noise subtractor exactly the reference it needs. Our mechanism is designed around this: ultraviolet actuation is invisible in RGB frames, so only the film's localized, subregion-patterned response appears in the visible channel, with no global artifact to cancel. Our experiment includes a projector arm to verify this reasoning empirically rather than assume it.

Why not just attack the software — backdoor the detector? The obfuscated-backdoor route of the Insidious Pipeline proves white-box undetectable, non-replicable backdoors exist, but by its own construction relies on heavyweight iO with extreme polynomial overhead, making it computationally intractable for real networks, and it requires code access. Physical presentation attacks target deployed, black-box verification systems the attacker cannot compile. These are complementary threat models, not substitutes — indeed, one natural sequel is the physical trigger interface sketched by Lou et al., whose heat-modulating backdoor trigger is the electronics-based cousin of what we propose.

Proposed Experiments and Decisive Checks

The evaluation proceeds in three stages. First, benchtop characterization of the film on flat coupons and then on the 3D replica: modulation depth in green-channel chrominance and in LWIR apparent temperature as a function of driving wavelength, intensity, and frequency; reversibility over extended cycling, benchmarked against the 70-plus cycles and two-second switching reported for the inverse-opal film and the one-minute UV shape recovery of the hydrogel. The go/no-go check for stage one is whether the film sustains a roughly 1 Hz oscillation with sufficient depth in at least the chrominance channel — if the host-guest kinetics prove too slow, we will report that as a negative result and pivot to thinner films or higher-intensity drivers, since the mechanism's speed, not the chemistry's existence, is the open variable.

Second, digital and hybrid evaluation. We extend the CDUPatch adapter to a time-series RGB-to-thermal mapping trained on our own paired recordings (no reported result found on Lacuna for any dynamic version of such an adapter), optimize the driving waveform against the CFrPPG feature extractor, and evaluate end-to-end spoof success on held-out detector checkpoints, including the material classifier of Liu et al. at its published HiFiMask operating point (2.4% ACER) and the intrinsic-image method at its 3DMAD point (0.0% EER).

Third, physical deployment: the driven mask worn by a mannequin headform and by a human volunteer holding it, under the indoor illumination conditions of the HKBU-MARsV2 protocol and the seven-device multi-sensor protocol of the HiFiMask challenge, with a handheld and a fixed thermal camera. Metrics are EER and ACER shifts relative to each detector's published operating point, the matched-versus-ablated waveform gap, and the LED-mask and projector comparison arms. The decisive check for the whole proposal is a curve: defender error as a function of attacker waveform fidelity, traced from the strong baseline toward chance. Wherever that curve saturates below 50%, the remaining separation is the defendability gap, and characterizing it — diffusion phase lag, harmonic content, inter-region delay mismatch — is the deliverable that turns an attack into a defense roadmap.

We close with the framing that makes this more than an attack paper. Christiano et al. proved that perfect indistinguishability makes defense impossible; the practical question for biometrics is how much indistinguishability real matter can buy, and at what cost. By building the strongest possible physical impostor and measuring exactly where it stops working, we give the liveness-detection community the adversarial boundary it currently lacks — the same service that physical adversarial patches rendered to object detection, but for the physiological channel that everyone still trusts.