A face morphing attack blends two people's photos into one passport portrait. Face recognition software then accepts that single document for both people. Current morphs are very hard to catch. Diffusion-based morphs — made with generators that build an image by repeatedly denoising random noise — already look realistic. Template inversion morphs fool face matchers most of the time: 3D Face Morph Generation Using Geometry-Aware Template Inversion reports a Mated Morph Presentation Match Rate of 90.70% against a matcher the attacker never saw. Every one of these generators optimizes for two targets: a 2D photo that looks real and a match-score embedding that sits close to both contributors. None of them is asked to produce a face whose three-dimensional anatomy agrees with a real person standing at the gate. That anatomy includes nose bridge depth, how deep the eye sockets are, and jaw curvature.
We propose a detector for the border-gate setting, where officers see both the passport photo and a live camera image of the traveler. The detector lifts both photos into a shared, pose-normalized 3D grid of small cubes (voxels) around the face, subtracts one grid from the other, and classifies the difference with 3D convolutions. The bet is that a morph and the accomplice who presents it disagree more in rigid 3D structure than in the features face matchers use. Two further properties should help. That structural disagreement is low-frequency, so it should survive the print-and-scan step of passport issuance better than pixel-level traces do. And a canonical 3D frame lets the detector compare the same anatomical point across two photos taken at different poses, which a flat 2D difference cannot do.
Core Claim
Differential morphing attack detection (D-MAD) compares a suspect document against a trusted live capture. It becomes more robust to unseen morph generators and to print-scan degradation when the comparison is made between anatomically registered 3D face volumes rather than between 1D identity embeddings or raw 2D images.
We state the information-theoretic caveat up front. A volume inferred from one photo contains no information that the photo lacks. The value of lifting is therefore an inductive bias: the 3D reconstruction prior forces comparisons into anatomical correspondence, and it separates geometry from texture and lighting. The second input, the live capture, is where genuinely new 3D evidence enters. It is a different view of a real, rigid head.
Relation to Prior Work
Two papers motivate the claim. Reflection Analysis for Face Morphing Attack Detection showed that landmark morphing distorts nose curvature enough to displace specular highlights. It predicted where highlights should fall using a trusted 3D scan and a hand-annotated cornea used as a light probe. We keep its physical intuition but drop the scanner and the manual annotation, and we test generators beyond landmark morphs. Neural Implicit Morphing of Face Images and the diffusion morphs of Leveraging Diffusion for Strong and High Quality Face Morphing Attacks and Greedy-DiM remove the 2D blending seams that detectors depend on. This is why we look for evidence below the texture level.
Other work constrains the design. Impact of Image Context for Single Deep Learning Face Morphing Attack Detection found that 2D detectors lean on face-contour boundaries and are sensitive to how the face is cropped. A volumetric detector must therefore not quietly reduce to a contour detector, and we test for this explicitly below.
The method comes from three sources. MoFaNeRF defines a continuous 3D face field conditioned on shape, appearance and expression codes. It was trained on the FaceScape and HeadSpace 3D face collections and used only for graphics, never for security. We reuse it as the lifting prior and sample its density and color onto a canonical voxel grid. The 3D-convolution designs of 3D U-Net and 3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes come from medical imaging, where they resolve structures that span many slices. We adopt them for a small face volume of roughly 64³ voxels, which keeps the cubic cost modest.
The closest detector is PoolAtnRes. It feeds signed differences of ArcFace identity embeddings through pooling, attention and residual blocks. We replace its 1D embedding difference with a registered volumetric difference.
Research Program
The pipeline has three stages.
Stage 1: lifting. An encoder is trained on bona fide faces only, using FaceScape and HeadSpace. It maps a single photo to the latent codes of a MoFaNeRF-style field. The field is sampled on a fixed grid in a canonical, expression-neutralized head frame. Each voxel stores density, color, and the density gradient, which serves as a surface normal. Pose and expression are factored out at this stage. That factoring is what makes a passport photo and a gate photo comparable voxel by voxel.
Stage 2: comparison. The enrolled volume and the live volume are combined as a signed difference and fed to a 3D residual classifier. The classifier also receives each volume's reconstruction residual: how badly the fitted field re-renders its own input photo. This residual carries the single-image cue that a morph may not correspond to any consistent head under one lighting setup.
Stage 3: controlled testbed. The detector is trained on real morph data, but first we build a testbed where 3D ground truth exists. We render frontal "passport" images of HeadSpace subjects and morph pairs of these renders using landmark, StyleGAN, diffusion and template-inversion generators. A different rendered view of one contributor serves as the "live" capture. This testbed answers the question the whole idea depends on, before any detector is trained: when a morph is lifted into 3D, does its geometry sit measurably far from the accomplice's true geometry? Or does the reconstruction prior pull the morph back toward a plausible face and erase the signal?
The main analysis measures where the 3D difference concentrates. It separates rigid regions (nose, orbits, jaw) from expression-sensitive regions, and it separates geometry channels from color channels. It also tests whether that signal drops less than 2D signals after print-scan.
Baselines
The naive reference point is plain face verification: we threshold the ArcFace cosine similarity between the enrolled and live images and treat low similarity as evidence of a morph. Any detector that cannot beat this adds nothing beyond the matcher already installed at the gate.
The strongest existing differential method is PoolAtnRes. It reports a mean D-EER (the detection equal error rate, where false alarms and missed attacks are equal) of 4.46% ± 3.06 under the intra-dataset protocol and 5.25% ± 0.68 under the inter-dataset protocol on the Multi-Morph Dataset (MMD). The paper notes weaker results on post-processed landmark morphs (MT #3), but the file gives no per-technique figure, so for that subset we record "no reported result found on Lacuna". We will retrain PoolAtnRes ourselves on identical splits with the same hyperparameter-search budget in GPU-hours as our model. Our method beats it only if the mean D-EER over MMD's nine techniques is lower and the gain is not confined to one technique. On the print-scan subset and the MT #3 subset, it must also do better at a fixed low attack-acceptance operating point, not only at the equal-error point.
As a single-image reference, we run frozen DINOv2 features with a one-class classifier, following Evaluating the Effectiveness of Attack-Agnostic Features for Morphing Attack Detection. That paper reports 8.85% D-EER on landmark print-scan morphs and under 1% on digital morphs using DNADet features. We also add DINOv2 difference features, fed to the same classifier head as ours, as a differential variant. This gives the foundation-model approach a fair shot at the same two-image setting.
Three ablations isolate the key ingredient. First, a geometry-only 2.5D control: the same lifting encoder is used, but the output is a canonical UV position map (per-pixel 3D coordinates on the unwrapped face) processed with 2D convolutions. This separates "canonical 3D geometry" from "volumetric processing." Second, a 3D Morphable Model (3DMM) shape-coefficient difference fed to a small multilayer perceptron. Third, our full model with the color channels removed, and separately with the live capture removed. These last two show whether the gain comes from geometry and from the differential comparison.
Why not just…?
Why not just use a frozen foundation model like DINOv2? These models are strong on digital morphs, but print-scan pushes them to 8.85% D-EER on landmark morphs in Evaluating the Effectiveness of Attack-Agnostic Features for Morphing Attack Detection. Their features are 2D appearance statistics, and printing distorts exactly those statistics. The differential DINOv2 baseline settles the question directly. If it matches our model on the print-scan subsets, then lifting adds nothing and we will say so.
Why not just improve 2D embedding differences, as in PoolAtnRes? ArcFace embeddings are trained to be invariant to much of the geometry we care about. Morph generators such as template inversion optimize the midpoint of two embeddings directly, which makes embedding differences a target the attacker already attacks. Head-to-head results on held-out generators, including Greedy-DiM and template inversion, will show whether this concern is real.
Why not just fit a 3DMM and compare shape coefficients? Linear 3DMMs cannot represent fine non-Lambertian structure (shine, reflections, fine surface detail that simple matte shading misses), as argued in MoFaNeRF. They also compress shape into a few dozen principal components, which may average away the local nose and orbit distortions seen in Reflection Analysis. This is a cheap and credible competitor, so it appears as an ablation rather than a straw man.
Why not just use a monocular depth map or UV position map with a 2D network? This is the most serious threat. If canonical geometry is what matters, a 2.5D map may capture nearly all of it at a fraction of the cost. The UV-map ablation answers this. If it comes within one D-EER point of the voxel model, we will adopt it and reframe the contribution as registered-geometry differential detection. The volumetric version keeps its place only if its advantage appears in regions a single surface map cannot represent well, such as self-occluded areas under pose change and density near hair and ears.
Proposed Experiments
The first check is the go/no-go test on the HeadSpace testbed. For each morph generator, we compare two distributions of lifted-geometry distance. One is the distance between the morph and its accomplice's second view. The other is the distance between a bona fide image and the same subject's second view. If the two distributions overlap heavily, the reconstruction prior is erasing the signal, and we stop or retrain the lifter with less shape regularization.
The main evaluation uses MMD. We run intra-dataset and leave-one-technique-out protocols over its nine techniques and report digital and print-scan results separately. We add FRGC-derived pairs and new FRLL morphs made with Greedy-DiM and template inversion, with FRGC uncontrolled-session images serving as live captures. Metrics are D-EER and the bona fide rejection rate (BPCER) at a fixed 1% attack acceptance rate (APCER). We also report generator strength as MMPMR so that improvements are not simply the result of weak attacks.
A contour-leakage test addresses the risk raised by the image-context study. We mask the face boundary and background in the input photos and check that the gain holds. This confirms the detector is not merely rediscovering boundary artifacts.
Finally, an adaptive-attacker test probes durability. We add a 3D-consistency loss, computed with our own lifter, to Greedy-DiM's step-wise guidance and measure how much D-EER recovers for the attacker. This shows whether the geometric cue lasts or is only a gap that attackers have not yet targeted.
References
- 3D Face Morph Generation Using Geometry-Aware Template Inversion
- Reflection Analysis for Face Morphing Attack Detection
- Neural Implicit Morphing of Face Images
- Leveraging Diffusion for Strong and High Quality Face Morphing Attacks
- Greedy-DiM: Greedy Algorithms for Unreasonably Effective Face Morphs
- Impact of Image Context for Single Deep Learning Face Morphing Attack Detection
- MoFaNeRF: Morphable Facial Neural Radiance Field
- 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation
- 3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes
- PoolAtnRes: Towards Generalisable Differential Morphing Attack Detection
- Evaluating the Effectiveness of Attack-Agnostic Features for Morphing Attack Detection