Liveness detectors decide whether a biometric sample comes from a live person or from a fake such as a printed photo, a replayed video, a silicone mold, or a mask. They are almost always trained one trait at a time and one institution at a time. Face data sits in one silo, iris data in another, and fingerprint data in a third. Privacy law keeps those silos apart, and each one is small. Federated learning, where institutions train a shared model by exchanging weight updates instead of raw data, has helped within a single trait. It has never been tried across traits. The usual objection is that a face network and an iris network have nothing in common to average.
This note proposes that they do share something: the physical medium of the attack. A paper print leaves halftone dots, ink bleed, and flattened reflectance whether the printed image is a face or an iris. Cast silicone has the same surface statistics on a fingertip as on a mask. We propose to federate only a small patch-level "artifact encoder" that looks at high-frequency residuals, where anatomy mostly disappears. Each client keeps its own trait-specific front end and decision head. A shared coarse vocabulary of attack media (print, screen, cast material, bona fide) ties the clients together. The central, falsifiable claim is that cross-trait federation helps a client exactly on the attack media it shares with other traits. Adding iris-print clients should improve face print detection on unseen datasets. It should do nothing for face screen replay, which no iris dataset contains.
Core Claim
Two federated systems bracket the existing work. Federated Face Anti-spoofing (FedPAD) trains identical ResNet-18 models on seven face datasets and averages their weights. It shows that attack diversity across clients transfers, since federating with 3D-mask datasets lets the global model catch mask attacks. DDP-FedFV splits a finger-vein recognizer into a shared domain-invariant extractor and a local domain-specific extractor to limit client drift. Both stay inside one trait.
The research file also contains a constraining result. Cross-Modality Generalization in Biometric Systems argues that face, iris, and vein features lie on incompatible structures, so naive parameter averaging across traits should cancel out. We take that seriously. We do not claim whole networks can be averaged across traits. We claim that one narrow component, a patch-scale encoder of residual texture, is trait-agnostic enough to share, because at that scale the dense spatial structure of a face and the annular texture of an iris are both mostly removed.
Supporting evidence comes from Deep Representations for Iris, Face, and Fingerprint Spoofing Attack Detection. That paper finds that one family of learned convolutional representations captures spoof artifacts well in all three traits. However, it trains separately per trait, so it never tests whether the representations transfer. The de-spoofing view of Modeling Spoof Noise by De-spoofing Diffusion supplies the conceptual input: an attack is a residual "spoof noise" layered on genuine structure. We adopt that view but approximate the residual with a cheap learned high-pass stem rather than a diffusion reconstruction, so edge clients can afford it.
What is new is not "federated learning plus multiple traits." It is a specific, testable hypothesis about which knowledge crosses traits, namely the attack medium. We pair it with an architecture that shares only the parts carrying that knowledge. The evaluation is designed to show gains where media overlap and no gains where they do not.
Research Program
Each client holds one trait from one source dataset. A client has three parts.
The first is a local stem. It crops the trait region and applies a small learnable high-pass convolution, which produces a residual image with low-frequency anatomy suppressed. The stem stays local because sensors and traits differ.
The second is the shared artifact encoder. It is a small CNN, a truncated ResNet-18 trunk, applied to randomly sampled 32×32 residual patches after per-patch contrast normalization. This is the only feature-extracting component whose weights are federated.
The third is a local pooling-and-decision head. It combines patch embeddings with attention pooling and outputs a live-versus-spoof score.
A second shared component is a small medium classifier on patch embeddings. It predicts a coarse label from a vocabulary common to all clients: bona fide, print, screen, or cast/molded material. Every client can assign these labels from its dataset's attack metadata. For example, iris prints in Warsaw and Biosec count as print, fingerprint silicone and gelatin count as cast material, and face masks in 3DMAD and HKBUMARsV2 count as cast material. The server averages the artifact encoder and the medium classifier. Stems and decision heads never leave the client.
Aggregation cannot be plain sample-weighted averaging. That would let the seven face clients swamp the iris and fingerprint clients, so we weight updates by trait before weighting by size within a trait. We also test a variant that weights clients by the overlap of their medium-label histograms, similar in spirit to how FedPWRR in DDP-FedFV weights clients by feature similarity.
The analysis program has three parts. First, it measures where transfer happens by breaking held-out performance down by attack medium. Second, it probes the shared encoder: can a linear probe trained on face-print patch embeddings detect iris-print patches, and does it fail on screen patches? Third, it checks the cancellation risk raised by the cross-modality paper. We compare the cosine similarity of face and iris client updates to the shared encoder against the similarity of updates to a whole-image encoder.
Baselines
The naive reference point is local training. Each client trains the full model on its own data alone, and we evaluate it on the held-out dataset of the same trait. FedPAD reports that the average single-dataset face baseline reaches 70.36% AUC. Every federated variant must clear this.
The main competitor is siloed same-trait federation, meaning FedPAD run separately for face, iris, and fingerprint. It uses identical ResNet-18 models, FedAvg, and the same leave-one-dataset-out protocol, with no cross-trait exchange. This is the strongest competitor on the question we ask, because it already captures all within-trait diversity, and it is the most likely reason a reviewer would call cross-trait sharing unnecessary. On face, FedPAD reports 76.51% AUC and 32.17% HTER (half total error rate, the average of false accept and false reject rates). The same paper reports 35.75% HTER for simple score fusion and 80.42% AUC for a centralized upper bound. For federated iris or fingerprint liveness detection we found no reported result on Lacuna, so we will establish those siloed numbers ourselves.
Fairness requires several controls. We reproduce FedPAD in our own code on the same splits. We give all methods the same number of communication rounds, local epochs, and hyperparameter search budget, with 20 configurations each, selected on source-domain validation data only. We also run our patch-residual architecture in siloed mode, so that architectural gains are not credited to cross-trait sharing. The proposal beats this baseline if cross-trait federation lowers held-out HTER on print-containing face targets, with confidence intervals over five seeds that exclude zero, while not raising HTER on screen-replay targets.
The strongest non-federated competitor is centralized pretraining followed by per-client adaptation with Optimal Transport-Guided Source-Free Adaptation for Face Anti-Spoofing (OTA). OTA ships frozen multi-centroid prototypes and shifts them toward a few target samples by optimal transport. It reports a host-prototype HTER of 5.29%, against 12.52% for GAC-FAS, on a face protocol that includes CelebA-Spoof. That protocol differs from ours, so the number is context and not a direct target. We will run OTA's training-free and geodesic-mixup variants with the same few-shot target budget we allow federated methods for optional personalization. The host is pretrained on pooled face data, which gives OTA a privacy advantage our setting forbids, so beating or matching it counts as strong.
Four ablations isolate the key ingredient. The first removes the shared medium classifier. The second replaces patch-scale residual input with whole-image input to the shared encoder while keeping everything else, which directly tests the cancellation concern. The third federates the whole network except the heads, a cross-trait version of the DDP-FedFV decoupling. The fourth shuffles medium labels across clients, which keeps the auxiliary loss but destroys its shared meaning.
Why not just…?
Why not just run a separate federation per trait? This is the real threat, and the proposal is designed so it can win. If siloed federation matches cross-trait federation on print-containing face targets, then the medium hypothesis is wrong for these datasets, and we will report that. The asymmetric prediction guards against a spurious win in which extra clients help through regularization alone. Such a win would show up as a gain on screen-replay targets too, and the medium hypothesis predicts no gain there.
Why not just average whole networks across traits with trait-specific heads? The cross-modality paper predicts that this cancels. DDP-FedFV shows that decoupling helps within a trait but offers no evidence across traits. The whole-image and DDP-style ablations test this directly. If either one matches the patch-residual design, then the patch restriction is unnecessary, and the contribution shrinks to the empirical fact that cross-trait sharing helps.
Why not just pretrain centrally and adapt with OTA? In this setting, pooling face, iris, and fingerprint data centrally is exactly what the data silos prevent. OTA is also one-directional, so a new attack medium seen at one client never reaches the others. Personalized Feature Translation for Expression Recognition shows that source-free adaptation degrades when targets fall outside source assumptions. In its case the failure was older subjects in BioVid, which is weak but relevant evidence. A decisive test is a new-medium experiment. We withhold all print attacks from the host and from face clients, and let only iris clients hold print attacks. The question is whether face print detection improves after federation, which OTA cannot do by construction.
Why not just fuse scores from per-trait detectors? Score fusion needs the same person to present several traits at test time, and our clients hold one trait each. FedPAD also reports that fusion is worse than federation within face, at 35.75% versus 32.17% HTER. Fusion also cannot move knowledge from a data-rich trait to a data-poor one, which is the benefit we are testing.
Proposed Experiments
The data are the seven face datasets used by FedPAD (OULU-NPU, CASIA-MFSD, Idiap Replay-Attack, MSU-MFSD, SiW, 3DMAD, HKBUMARsV2). We add the iris sets Biosec, Warsaw, and MobBIOfake, and the fingerprint sets Biometrika, CrossMatch, Italdata, and Swipe, as used in the deep-representations paper. Each dataset is one client.
The main protocol holds out one face dataset at a time and trains in two conditions: federation over the remaining face clients only, and federation that also includes all iris and fingerprint clients. We report HTER and AUC overall, then split by attack medium. The medium split is essential. The predicted pattern is a gain on print attacks and roughly zero change on screen replay. Face-mask targets, 3DMAD and HKBUMARsV2 held out, receive fingerprint cast-material clients; we expect a smaller or absent gain there, because mask resin and fingerprint gelatin differ more than two paper prints do. We repeat the protocol in the reverse direction, holding out one iris dataset and adding face clients. Iris is data-poor, so if the hypothesis holds, the reverse gain should be larger.
Three decisive checks follow. The first is the new-medium experiment described above. The second is the linear-probe transfer test on the shared encoder, trained on face-print patches and evaluated on iris-print and face-screen patches. The third is the update-similarity measurement comparing patch-scale and whole-image shared encoders.
All results are averaged over five seeds, with bootstrap confidence intervals over held-out samples. We also report communication cost per round, which is small because only the artifact encoder and medium classifier are transmitted, and local compute. DDP-FedFV notes that decoupling raises edge overhead, so we report it rather than assume it away.
The hypothesis is confirmed if print targets gain and screen targets do not. It is refuted if gains are uniform across media or absent. Either outcome answers a question the field currently settles by assumption.
References
- Federated Face Anti-spoofing
- DDP-FedFV: A Dual-Decoupling Personalized Federated Learning Framework for Finger Vein Recognition
- Relativistic Discriminator: A One-Class Classifier for Generalized Iris Presentation Attack Detection
- Deep Representations for Iris, Face, and Fingerprint Spoofing Detection
- Modeling Spoof Noise by De-spoofing Diffusion and its Application in Face Anti-spoofing
- Optimal Transport-Guided Source-Free Adaptation for Face Anti-Spoofing
- Personalized Feature Translation for Expression Recognition: An Efficient Source-Free Domain Adaptation Method