Challenge–Response Liveness from Plasmon-Tuned Lanthanide Tags

Deepfake defenses today live almost entirely inside files. Detectors look for statistical traces in pixels, and watermarks are hidden in pixels or in generator weights. Both lose their footing once the question becomes physical: is this camera looking at a real person right now, or at a screen, a print, or a rendered avatar? We propose a small wearable or stick-on optical tag. The verifier flashes it with a short, randomly timed and randomly coloured light pattern, and the tag answers with a glow that fades over milliseconds. An ordinary phone camera can read that fade. The tag's answer depends on the challenge in a way fixed by nanometre-scale disorder in the tag itself. An attacker who has only video of past answers cannot synthesize the right answer to a new challenge in real time, even with a perfect face generator.

The glow comes from lanthanide ions such as europium (Eu³⁺). These metals barely absorb light directly, so they are "sensitized": an organic molecule bound to the metal (the antenna, or ligand) absorbs the light and hands the energy to the metal ion, which then emits narrow red light lasting milliseconds. We place these complexes near gold nanoparticles separated by a thin insulating spacer. At the right distance, the metal's concentrated near-field shortens and brightens the emission. The visible-light-driven plasmon can also inject energetic ("hot") electrons that quench it. Spacer thicknesses vary randomly at the nanometre scale across the tag, so each spot on the tag ends up with its own lifetime and its own dependence on challenge colour. The tag is an unclonable, time-resolved, wavelength-addressed fingerprint. The ML contribution is the verifier and an adversarial evaluation against the generators in the research file.

Core Claim

A tag whose emission decay time and challenge-colour dependence are set by uncontrolled nanometre spacer disorder gives a liveness signal with three properties. A commodity rolling-shutter camera can read it. It resists photographic copying and resists replay or generation from recorded footage. And it is neither a static pattern nor a uniform material, which is the step beyond existing physical anti-counterfeiting work.

The camera point needs explanation. A rolling-shutter sensor reads its rows sequentially, roughly 10 µs apart, so a single frame contains a time-resolved trace of a millisecond decay. No specialised photodetector is needed.

The chemistry and physics make the timescales plausible. Sensitized emitters routinely reach millisecond lifetimes; the antenna phosphorescence in 4D Assembly of Time‐dependent Lanthanide Supramolecular Multicolor Phosphorescence reaches 7.4 ms. Nanogaps can accelerate emission by more than 10³ (Qian et al.). The steep distance dependence of enhancement versus quenching (Distance-Dependent Plasmon-Enhanced Fluorescence of Submonolayer Rhodamine 6G) means that spacer-thickness disorder of a few nanometres should translate into large lifetime variation.

Relation to Prior Work

Three groups of prior work bear on this proposal, each in a different role.

The first group motivates it. Active digital defenses stop at the file boundary. Pluggable Watermarking of Deepfake Models embeds error-corrected codes in decoder weights and needs cooperative generator owners. CMUA-Watermark perturbs source photos before upload, against models it anticipates. DIP-Watermark hides a recoverable identity code in feature space. None of the three can tell a live scene from a recaptured screen, and none defends against an attacker who trains a custom avatar from scraped video. Passive detection is weak on refined fakes: the average detector AUC falls to 56.9% on Celeb-DF.

The second group supplies the method. Lanthanide sensing work shows that the environment can modulate ligand-to-metal energy transfer and that the modulation can be read optically (Eu-MOF ratiometric sensor; Tb(III) framework for Ba²⁺). The 4D clay system uses time-evolving emission for encryption. Ag–SiO₂ core–shell solar cells and the R6G/PMMA study show that a dielectric shell decides whether a metal enhances or quenches nearby emitters. Bare Ag drops PCE to 1.89%, while silica-coated Ag raises it to 3.96%.

The third group constrains the design. Physical unclonable labels already exist (diamond microparticle PUFs; multimodal dynamic diamond labels). They are read as spatial patterns, and a spatial pattern can be photographed and re-rendered. The 4D clay material changes over an hour, which is far too slow to answer a per-session challenge. Our proposal therefore differs from them along one axis: the answer must depend on a fresh challenge on a millisecond timescale.

The hot-electron literature supplies a caution. Extraction efficiencies are typically below 1–10%, so the hot-electron quenching channel may be too weak to measure. We treat it as a hypothesis to test, not a load-bearing assumption.

Research Program

The work has three parts.

The tag. We will fabricate Eu³⁺ complexes with a phenanthroline-type antenna, the same ligand family as in Eu/Tb(BPA)₃phen composites. They go over a film of Au nanoparticles coated with a deliberately non-uniform silica or PMMA spacer of nominal thickness 2–15 nm, which spans the regime that the R6G study maps from quenching to enhancement. The tag sits in a rigid oxygen-excluding matrix, because triplet antennas are quenched by O₂.

The two challenge channels. Near-UV excitation drives the antenna. Visible excitation near the gold resonance (about 530–580 nm, per the R6G study's resonance shift) drives the plasmon. The tag's response is a function of challenge colour, timing and tag position. Writing $\tau(x)$ for the local decay time at tag position $x$ and $R(x)=I_{\mathrm{UV+vis}}(x)/I_{\mathrm{UV}}(x)$ for the change in emission when visible light is added, a tag instance is the joint map $\{\tau(x), R(x)\}$. A challenge is a random sequence of pulse times and colours. The expected response follows from convolving that sequence with the local decay. We will characterise these maps with a time-correlated photon-counting setup, which gives the reference readout, and with phone cameras. The phone pipeline estimates per-region decay curves from rolling-shutter row timing, and we will compare its estimates against the photon-counting ones.

The verifier. A small model takes the observed frames and the issued challenge. It predicts whether the response is consistent with the enrolled tag maps, and is trained only on genuine responses plus generic camera noise. We will compare it against a purely analytic matched-filter verifier. If the matched filter is as good, we use it, since it is easier to audit.

Baselines

The naive reference is software-only liveness with no tag. CMUA-Watermark reports that unprotected deepfakes pass the HyperFAS liveness check at a rate of 66.3%. We will run HyperFAS-style face anti-spoofing on our own attack set so that both arms see identical videos.

The strongest passive competitor is a frame-level deepfake detector. DSP-FWA reports 87.4% average AUC in the Celeb-DF evaluation. We will train it with the same data budget on our genuine and attack clips and report AUC on held-out generators.

The strongest active competitor is weight-level watermarking. Pluggable Watermarking reports detection accuracy above 94%, staying above 90% under JPEG Q=60. That comparison is only fair when the attacker uses a cooperating, watermarked generator, so we will report it in that setting and separately for a custom, unwatermarked generator, where it gives no signal. Neither paper reports a result for screen recapture or for live physical liveness; no reported result was found on Lacuna for any defense under analog recapture.

Two ablations isolate the key ingredient. The first is a static tag: the same lanthanide film read as a spatial pattern only, which is our analogue of the diamond PUF labels. The second is a uniform-lifetime tag: the same chemistry with no gold, which has a challenge-dependent response but no unclonable variation.

For the tag to count as beating the baselines, three conditions must all hold at a 1% false-reject rate. Its false-accept rate must be below that of the best software baseline on every attack class. It must be below the static-tag ablation on the generative-replay attack. And it must be below the uniform-tag ablation on the cloning attack.

Why not just…?

Why not just sign images at capture time? A cryptographically signing camera proves which device took a photo, not what was in front of the lens. A signed recapture of a screen playing a GaussianAvatars render at 187 FPS is still signed. Signing is complementary, not a substitute. Our screen-recapture attack, run through a signing pipeline, would show this directly.

Why not just use a static fluorescent or diamond-particle tag? This is the most serious alternative. A static tag can be filmed once and re-rendered. Modern avatar models such as TalkingGaussian reproduce appearance faithfully from about 30 minutes of training, so the tag would simply become part of the learned appearance. The static-tag ablation measures exactly this. If static tags survive generative replay as well as dynamic ones, the dynamic design is unnecessary and we will say so.

Why not just use lanthanide lifetime alone, without plasmonics? A uniform-lifetime tag already answers challenges correctly. Its weakness is that anyone with the same recipe can make a working copy. The gold and spacer disorder is what makes each tag unique. The cloning experiment, in which we fabricate a second tag from the same recipe and test it against the first tag's enrolment, decides whether that disorder is worth the fabrication cost.

Why not just improve software liveness detection? Celeb-DF shows that detectors trained on one family of artifacts collapse on refined fakes. Generator quality keeps rising: OmniSync reports 97.4% generation success on the AIGC-LipSync benchmark. A physical challenge whose answer is unknown before the session does not depend on spotting generator artifacts.

Decisive Experiments

The physics check comes first. The tag works only if spacer disorder produces a lifetime spread larger than the camera's lifetime-estimation error. We will measure both: the spread with photon counting, and the error by comparing phone rolling-shutter estimates with photon-counting estimates on the same spots. Separately, we will test whether adding visible light changes $R(x)$ beyond the noise floor. If it does not, hot-electron quenching is too weak at these efficiencies, and we drop that channel and keep the Purcell-only design.

The security evaluation then uses about 30 consenting participants wearing tags. There are four attack classes:

  • Screen replay: genuine past sessions are replayed on a high-refresh display.
  • Generative replay: GaussianAvatars, TalkingGaussian, StyleSwap and OmniSync are trained on each participant's recorded footage. For this class we also train an adaptive attacker: a video model conditioned on the issued challenge and on all previously observed challenge–response pairs, given a fixed compute budget.
  • Cloned tags: copies are fabricated from the same recipe.
  • Re-encoding: genuine sessions are passed through social-media-style compression to measure false rejects.

The adaptive generative attacker is the critical one. The claim holds only if the attacker's false-accept rate stays near chance as the number of observed sessions grows. If the attacker learns the tag's response function from a few hundred sessions, the tag is merely an obscure static secret. In that case the honest conclusion is that the tag needs periodic re-enrolment, and the paper would report how many sessions the tag's response maps can withstand before an attacker succeeds.

References

  1. 4D Assembly of Time‐dependent Lanthanide Supramolecular Multicolor Phosphorescence for Encryption and Visual Sensing
  2. Spontaneous emission in micro- or nanophotonic structures
  3. Distance-Dependent Plasmon-Enhanced Fluorescence of Submonolayer Rhodamine 6G by Gold Nanoparticles
  4. Pluggable Watermarking of Deepfake Models for Deepfake Detection
  5. CMUA-Watermark: A Cross-Model Universal Adversarial Watermark for Combating Deepfakes
  6. DIP-Watermark: A Double Identity Protection Method Based on Robust Adversarial Watermark
  7. Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics
  8. Eu-MOF-Based Fluorescent Ratiometric Sensor by Detecting 3,4,5-Trihydroxybenzoic for Fingerprint Visualization on Porous Objects
  9. A Fluorine-Functionalized Tb(III)–Organic Framework for Ba2+ Detection
  10. Enhanced efficiency of polymer solar cells by incorporated Ag–SiO2core–shell nanoparticles in the active layer
  11. Manufacturing unclonable anti-counterfeiting labels using robust diamond microparticles on heterogeneous substrates
  12. Multimodal dynamic and unclonable anti-counterfeiting using robust diamond microparticles on heterogeneous substrate
  13. Synthesis and characterization of the novel color-tunable Eu/Tb(BPA)3phen composites
  14. GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians
  15. TalkingGaussian: Structure-Persistent 3D Talking Head Synthesis via Gaussian Splatting
  16. OmniSync: Towards Universal Lip Synchronization via Diffusion Transformers
  17. StyleSwap: Style-Based Generator Empowers Robust Face Swapping