Technology

Seven Layers. Geometric Verification. One Verdict.

Someone stole your mesh, retextured it, and re-uploaded it as their own. Shield finds it — and verifies it — even after rotation, scaling, retopology, or texture swaps.

Asset
Parse
7 Fingerprints
Compare
Verify
Verdict

Existing tools compare thumbnails.

We fingerprint and verify geometry.

Live Detection Spectrum

3D Shape
98%
Surface
93%
Texture
87%
Mesh
96%
Visual
94%
Topology
97%
Normals
91%
Clone Detected6 of 7 layers match

Evasion Resistance

What doesn't work.

Shield fingerprints the actual 3D geometry. These evasion techniques have no effect.

Retexturing

Different paint, same shape. Shield sees through it.

Polygon reduction

Fewer polygons, same distance distribution.

Retopology

New edge flow, same spectral signature.

Rotation & scaling

Every fingerprint is transformation-invariant.

Mirroring

Mirror images produce identical fingerprints.

Removing parts

Component stripping catches geometry additions.

Re-uploading

Fingerprints match across all accounts and listings.

Name changes

The mesh is the fingerprint, not the metadata.

Detection Layers

Seven independent analyses. Every scan.

Each layer targets a different evasion technique. Together, they identify clone candidates that move to geometric verification.

Layer 01

Visual

Re-colored copies · resized thumbnails

Two independent perceptual hashes capture visual appearance at the frequency level. Survives color shifts, crops, and resolution changes.

64-bitdual hash

Layer 02

Mesh

Direct mesh rips · unmodified geometry

Vertex count, triangle count, bounding box ratios, surface area. Parsed directly from Roblox’s binary mesh format. Eliminates 90% of non-matches instantly.

V1–V7all formats

Layer 03

3D Shape

Rotation · scaling · mirroring

Surface-area-weighted sampling creates a shape signature that’s mathematically invariant to rotation, scale, and translation. Enhanced with angle distribution analysis.

10Ksamples · 128 bins

Layer 04

Texture

Texture theft · material swaps

Frequency analysis of raw texture data catches “same shape, different skin” and “different shape, same skin” attacks independently.

64-bitDCT hash

Layer 05

Topology

Retopology · remeshing · subdivision

Analyzes intrinsic geometry that survives complete mesh rebuilds. Two meshes with the same shape produce nearly identical signatures even with different polygon structure.

0eigenvalues

Layer 06

Surface

Smoothing · faceting · detail changes

Captures how the surface bends at every vertex. Distinguishes items that share a silhouette but differ in surface detail.

0bins · angle deficit

Layer 07

Normals

Shape differentiation · orientation shifts

Surface orientation distribution — captures how mesh faces are directed in space. Discriminates compact objects that share similar size but have different shapes.

0bins · NDH

Geometric Verification

Statistical similarity is never enough.

After the seven detection layers identify candidates, Shield performs a direct geometric comparison on the actual mesh data to confirm identity. Two items that share similar statistical properties but are geometrically different will never be classified as clones.

Stage 1

Candidate Discovery

Seven detection layers analyze visual appearance, mesh structure, shape distributions, texture data, topology, surface curvature, and normals. This surfaces potential matches from the entire database.

Stage 2

Geometric Verification

Candidates undergo direct geometric comparison — Shield compares actual mesh geometry to confirm whether two items are genuinely identical. This eliminates false positives from items that happen to share statistical properties.

Visual Agreement

Disagreement Cap

Items that look visually different can't receive a Clone verdict — regardless of geometric similarity. This catches common-shape coincidences where two unrelated items share a geometry class but are clearly distinct.

Clone Detected requires geometric verification

Shield doesn't just find similar items — it verifies actual geometric identity. No item is ever classified as a clone based on statistical similarity alone. The geometric verification step is mandatory for every Clone Detected and Mesh Theft verdict.

Scoring Engine

Context-aware scoring. Not fixed weights.

A patch and a character aren't scored the same way. The engine classifies each item's morphology — flat, elongated, compact, complex — and adapts layer weights accordingly.

Cross-category intelligence

A hat and a necklace can share similar geometry without one being a copy of the other. Matches across different item categories are automatically capped at Review — never escalated to Clone Detected.

Verification gating

Statistical similarity alone is never enough for a Clone verdict. Every candidate must pass geometric verification — a direct comparison of actual mesh geometry. Without verified identity, the match is capped at Review.

What We Catch

Full Clone

Visual and structural layers agree, confirmed by geometric verification. Strong evidence of direct duplication with verified mesh identity.

Mesh Theft

Structural match with geometric verification, visual mismatch. Geometry stolen and retextured. Verified mesh identity confirms the theft.

Texture Theft

Visual match, structural mismatch. Same texture applied to a different mesh.

The Engine

Compiled Rust. Unreadable in production.

Written in Rust, compiled to WebAssembly, deployed to Cloudflare's edge network. Stripped, optimized, and running at near-native speed — everywhere.

Rust

Source language

WASM

Compile target

0

Automated tests

<3s

7 layers + verification

3D Evidence

See the geometry. Compare the proof.

Interactive 3D previews built from the same mesh data Shield fingerprints.

Every scanned mesh is simplified to ~5,000 vertices for preview. That simplified geometry is stored and rendered on demand — solid dark faces with colored wireframe edges, drag to orbit, scroll to zoom.

When a match is found, both meshes appear side-by-side with synced camera controls. Drag one viewport and both rotate together — your item in blue, the suspected clone in red. Visual proof for DMCA evidence that a thumbnail alone can't provide.

~5K

Preview vertices

R3F

React Three Fiber

0 KB

Initial bundle cost

0fps

Synced comparison

[1] Perceptual hashing — PhotoDNA (Microsoft), Content ID (YouTube), PDQ (Meta).

[2] Osada et al. (2002). Shape Distributions. ACM Transactions on Graphics, 21(4).

[3] Reuter, Wolter, Peinecke (2006). Laplace-Beltrami spectra as Shape-DNA. CAD, 38(4).

[4] Texture perceptual hashing via DCT frequency analysis.

[5] A3 angle distributions — triplet angle sampling for shape comparison.

[6] Mean curvature distributions — angle deficit curvature histograms.

[7] Normal distribution histograms — surface orientation analysis for shape discrimination.

fingerprints indexed and counting

See it in action.

Run a scan on any Roblox UGC item. Or register your portfolio and let the engine monitor the catalog for you.