/ modules / meshguard
module · AI neural watermark · mesh provenance

MeshGuard.

AI neural watermarking embedded into the mesh geometry itself — not into a sidecar file, not into a STL header, not into a metadata tag. The watermark survives slicing, scaling, infill choice, and the eighteen months of handling we measured in the field. Detection runs offline in eleven milliseconds on commodity hardware. False-positive rate, at the deployed decision threshold, sits below 10-9.

$problem

Why watermarks on meshes are hard.

The naive approach — embed a payload in mesh metadata — dies on the first slicer pass; slicers strip metadata, re-triangulate, decimate, repair. The slightly less naive approach — embed in vertex coordinates — survives slicing but produces detectable surface artifacts at the precision required for survival.

MeshGuard takes a different approach: a small neural network learns a mesh-level perturbation that is below the perceptual and metrological detection threshold, yet survives the standard slicer pipeline because it is encoded in topology-stable features (curvature spectrum, geodesic landmark spacing) rather than in raw vertex coordinates. The detector is a sibling network trained on the same feature space.

$ai-detector

The AI model is the product core.

MeshGuard uses paired AI models: an embedder that learns topology-stable perturbations and a detector that reads mesh or point-cloud features after manufacturing transformations. The model output is always paired with a signed payload and customer key binding.

embedderGenerates watermark perturbations below the customer's dimensional tolerance budget.
detectorScores mesh files, slicer outputs, or scanned point clouds against a customer payload.
triage modelRanks suspicious parts by confidence, supplier history, part class, and mismatch severity.
$algorithm

Algorithm sketch.

  1. Feature extraction. The embedder computes a topology-stable feature spectrum on the input mesh — curvature distribution, geodesic landmark spacing, surface normal entropy.
  2. Payload encoding. A 128-bit payload (customer identifier + part identifier + nonce) is expanded into a perturbation vector in feature space, signed by the customer's MeshGuard key.
  3. Perturbation embedding. The embedder back-projects the perturbation onto vertex positions, bounded by a configurable perceptual budget. The default budget is 12 µm maximum vertex displacement — below the dimensional tolerance of every printer process we have certified for.
  4. Mesh emission. The watermarked mesh is emitted in the customer's preferred format (STL, 3MF, OBJ). The watermark is in the geometry, not the file format.
  5. Detection. The detector accepts a mesh (or a point cloud sampled from a physical part) and produces a confidence score against a customer key. Scores above the decision threshold are accepted; below, rejected.

The detector runs on the same feature spectrum, which is what makes the watermark survive: a slicer cannot reasonably modify the curvature spectrum without producing a part that looks different to a human.

$survival

What it survives.

The field measurement programme (advisory 3DC-2026-04-A2) measured watermark survival across 41 sites, 11 material classes, 7 slicer builds, and 18 months of physical handling. Headline survival rates from the programme below.

operationsurvival ratenotes
slicer pass (Prusa / Cura / Orca / Bambu / Materialise)100.00%n = 41,184 slices, no failures
uniform scaling 0.5× to 2.0×99.998%1 failure at 2.0× on a coarse mesh, recovered after re-mesh
infill change (10% ↔ 100%)100.00%watermark is in the shell topology, infill is irrelevant
shell thickness change (1–6 perimeters)100.00%as above
orientation change100.00%feature spectrum is orientation-invariant by construction
FDM print + cooled handling (PLA, PETG, ASA)99.94%2 failures across 3,180 parts; failures correlated with extreme warping
SLS print + post-process (PA12, PA11, glass-filled PA12)99.86%4 failures across 2,840 parts; investigated, traced to surface bead damage
SLA print + UV cure99.71%residual cure shrinkage near vertical features; 0.5% across cohort
metal DED + machining finish97.40%finishing removes some surface; threshold increased for metal parts
18 months physical handling (cohort sample)99.62%n = 1,580 parts tested, parts returned from customer use

The numbers above are signed against the field-study programme key. The full report (advisory 3DC-2026-04-A2) carries the per-material breakdown, the regression analysis on warping, and the four edge-case failures we did not recover.

$fpr

False-positive rate.

What we publish is the FPR at the deployed decision threshold. The full ROC curve from the field study is in the report.

thresholdfalse-positive ratetrue-positive raterecommendation
0.50≈ 10-399.98%too lenient; rejected
0.85≈ 10-799.91%acceptable for non-regulated
0.93< 10-999.84%deployed default
0.97≈ 10-1299.62%recommended for high-stakes contested claims

The deployed default sits at 0.93. Customers with contested-claim use cases (e.g. legal evidence in counterfeit litigation) typically run a parallel scan at 0.97 and rely on the lower-threshold scan for ambient detection.

$detection

Detection in production.

Mesh-form detection.

The detector accepts a mesh file directly (STL, 3MF, OBJ, PLY). Useful when an OEM is checking a third-party CAD file uploaded to a licensed-manufacture portal. Runs at 11 ms median on a single CPU core for a typical 800k-triangle aerospace bracket.

Physical-part detection.

A point cloud sampled from a physical part (structured light scan, photogrammetry, CT for opaque parts) is sufficient. The detector reconstructs the feature spectrum from the point cloud. Detection on a 1.2 million-point scan runs in approximately 40 ms.

Both detection modes are offline. The detector binary is approximately 18 MB and ships with the SDK; the model weights are signed and pinned to a customer's MeshGuard key cohort. There is no telemetry to us during detection; we do not know what you scan or when.

$use-cases

Where MeshGuard earns its keep.

Licensed third-party manufacture.
OEMs that license out-of-warranty replacement printing keep track of which third parties manufacture under licence. MeshGuard embeds the licensee identifier in every shipped CAD; physical parts can be traced back to the licensee even after a year in service.
Counterfeit detection in spare-parts logistics.
Industrial spare-parts OEMs scan inbound parts at warehouse intake. Watermark match means a legitimate licensed copy; watermark absence triggers a deeper inspection. The cost per scan is below the cost of one call-centre minute on a returned counterfeit.
Provenance evidence in litigation.
Two ongoing IP-litigation cases (under NDA) use MeshGuard scans as part of their evidence chain. Detection at the 0.97 threshold provides confidence at the standard required.
Regulatory traceability.
Medical implants printed at multiple sites under the same OEM brand can be traced to their print site. The watermark resolves to a site-and-printer identifier that the OEM's quality system reconciles against its TwinCert records.
$limits

What MeshGuard does not do.

Honest limits, surfaced rather than buried.

  • It does not stop counterfeit production. A determined counterfeiter can manufacture without the watermark; what MeshGuard does is make the counterfeit identifiable as such after manufacture.
  • It does not survive arbitrary mesh remodeling. A counterfeiter who manually remodels a part in a CAD system (rather than printing your CAD file) produces a fresh mesh with no watermark. MeshGuard defends against the high-volume scenario, not the artisan one.
  • It does not survive aggressive surface-finishing on metal. Heavy machining removes the surface features the watermark sits in. Metal parts intended for heavy post-process should rely on TwinCert for provenance, not MeshGuard alone.
  • It does not perform cryptographic verification. MeshGuard tells you a part originated from your CAD; it does not tell you the part was actually made by an authorised printer. For that, you want Vault3D + TwinCert.
$integration

Integration shape.

For OEMs distributing CAD to third parties.

The embedder runs at the OEM's distribution point. A typical pipeline: licensee requests CAD via portal, portal calls MeshGuard embedder with the licensee identifier as payload, embedder emits a watermarked mesh, the watermarked mesh is what the licensee downloads. The OEM keeps the source mesh; the licensee gets a per-licensee variant.

For OEMs operating their own print fleet.

The embedder runs at the slicer entry — before any tool-path generation. The payload is per-printer, per-batch, or per-customer order, depending on the OEM's traceability policy. The watermarked mesh feeds the slicer normally; the resulting part is identifiable to its print event.

For verification at the receiving site.

The detector runs locally on the verifier's hardware. Mesh-form for digital files, point-cloud for physical parts. The customer console exposes a one-click scan flow; the SDK exposes the same operation for high-throughput pipelines.