Adversarial attacks on 3D deep networks almost always move coordinates: they nudge thousands of points or vertices by small amounts until the classifier flips. A parallel literature in shape matching knows something the attack literature has ignored: the connectivity of a surface, not its coordinates, controls its global frequency content. When a scanned hand touches a hip, or two fingers fuse, geodesic distances short-circuit across the artificial bridge and the entire eigenbasis of the Laplace operator — the shape's intrinsic Fourier basis — reorders and scrambles across the whole object. In this proposal I turn that fragility into an attack: instead of perturbing spectral coefficients with gradients, we add or remove a single connection between two nearby-in-space but far-along-the-surface points, and let the shape's own global spectral structure amplify a millimeter-scale edit into a model-breaking perturbation.
The claim has three parts. First, a one-edge edit is a closed-form spectral attack: adding an edge of weight $w$ between vertices $a$ and $b$ to a graph or mesh Laplacian $L$ shifts eigenvalue $\lambda_i$ by $\delta\lambda_i \approx w\,(u_i(a) - u_i(b))^2$, where $u_i$ is the $i$-th eigenfunction — no eigendecomposition per attack step, no optimization loop, and maximal damage exactly where the bridge short-circuits points that the eigenfunctions distinguish most. Second, because it acts on connectivity rather than coordinates, the attack is nearly free in the perturbation-size metrics the field uses (Chamfer and Hausdorff distance, number of moved points), and it looks like an ordinary scanning artifact rather than an adversarial edit. Third, it is representation-agnostic by construction: meshes carry explicit edges, point clouds carry $k$-nearest-neighbor edges, and the same rank-one update formula governs both, so an attack designed on one discretization transfers to another without the expensive inverse-spectral recovery that earlier cross-representation attacks required.
What the Literature Leaves Open
Three lines of work border this idea, and none occupies it. Spectral adversarial attacks are mature: the Graph Spectral Domain Attack perturbs graph-Fourier coefficients of point clouds and achieves a 100% success rate in targeted and untargeted settings against PointNet, PointNet++, and DGCNN on ModelNet40 while evading Statistical Outlier Removal and Simple Random Sampling defenses Exploring the Devil in Graph Spectral Domain for 3D Point Cloud Attacks; its extension GSDA++ parameterizes the attack as a learnable polynomial spectral filter and holds a 100% success rate against five architectures including PointTransformer and PointMLP Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing; SAGA perturbs low-frequency Laplace–Beltrami coefficients on meshes and exceeds 99% success forcing target reconstructions on CoMA, while a detector that catches vertex-based attacks at 98% accuracy drops to roughly 53% — near chance — against it SAGA: Spectral Adversarial Geometric Attack on 3D Meshes. All of these perturb coefficients given a fixed basis, assume clean topology, and pay an $O(n^3)$ eigendecomposition per shape.
The closest existing work to ours is the universal spectral attack of Rampini et al., which learns a multiplicative perturbation on the first $k$ Laplace–Beltrami eigenvalues and recovers physical coordinates through a band-limited "Shape-from-Spectrum" optimization, transferring across poses and even from meshes to point clouds — a perturbation learned on 15 shapes fooled PointNet on unseen CoMA test shapes with 80.8% success Universal Spectral Adversarial Attacks for Deformable Shapes. That paper motivates ours: it proves spectra are the right cross-representation attack surface, but it treats eigenvalues as a stable invariant, edits them through heavy optimization, and assumes watertight manifolds. We attack the basis through topology instead: one physical edit, no inverse problem, and — critically — we predict and exploit the fact that real shapes with fused limbs or holes violate its clean-manifold assumption.
The third border is the shape-matching literature, which documents the fragility we weaponize but never uses it adversarially. Localized bases were invented precisely because a single topological contact corrupts global eigenfunctions: multiscale manifold wavelets confine the damage locally Efficient Deformable Shape Correspondence via Multiscale Spectral Manifold Wavelets Preservation, $k$-hop neighborhood matching reaches 94.9% accuracy on the topologically noisy TOPKIDS benchmark where global spectral methods fail HOPE: Shape Matching Via Aligning Different K-hop Neighbourhoods, and differentiable coupling with test-time Dirichlet regularization reaches a geodesic error of 9.2 on TOPKIDS against 23.4 for the next best unsupervised method Unsupervised Learning of Robust Spectral Shape Matching. CoE extends robust matching to raw point clouds, whose $k$NN connectivity cannot distinguish a true surface neighbor from a topological short-circuit — a point that becomes, in our setting, the attack's delivery mechanism rather than a defect CoE: Deep Coupled Embedding for Non-Rigid Point Cloud Correspondences.
Core Claim and Mechanism
The proposal's central claim: minimal connectivity edits are a strictly cheaper and more transferable path to global spectral corruption than coefficient-level perturbation, and they simultaneously expose a shared failure mode across heterogeneous 3D representations — explicit meshes, $k$NN point-cloud graphs, and the tokenizers of unified models.
The mechanism is elementary spectral perturbation theory, which to our knowledge no 3D attack paper has used. Write the graph or cotangent Laplacian as $L$ with eigenpairs $(\lambda_i, u_i)$. Adding one edge of weight $w$ between vertices $a$ and $b$ changes the Laplacian by a rank-one matrix,
$$\Delta L = w\,(e_a - e_b)(e_a - e_b)^\top, \qquad \delta\lambda_i \approx w\,\big(u_i(a) - u_i(b)\big)^2,$$
where $e_a, e_b$ are indicator vectors. The formula says everything the attack needs. A bridge between two points that the low-frequency eigenfunctions barely distinguish (nearby along the surface) does almost nothing; a bridge between points that are close in 3D but far along the surface — palm to hip, finger to finger — moves every low eigenvalue substantially, because smooth eigenfunctions vary greatly across such pairs. One can rank candidate bridges by predicted total spectral damage $\sum_i \delta\lambda_i$ without re-decomposing the Laplacian, using the clean shape's eigenvectors once, or even replacing them with cheap surrogate descriptors (heat-kernel or geodesic-signature differences). Removing an edge is the same formula with opposite sign, and on point clouds a displaced point simultaneously deletes and creates $k$NN edges, each a rank-one term. The attack is therefore: pick the highest-ranked bridge site, fuse (or move $m \le 10$ of 1024 points to create) the connection, and stop. This is the whole attack — a contrast with GSDA++'s per-shape eigendecomposition plus iterative optimization Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing and with Rampini et al.'s Shape-from-Spectrum inverse problem Universal Spectral Adversarial Attacks for Deformable Shapes.
A note on a real contradiction in the literature: SAGA argues stealth requires perturbing low frequencies, while GSDA++ argues stealth requires preserving low frequencies and perturbing high ones, and the two imply opposite defenses SAGA: Spectral Adversarial Geometric Attack on 3D Meshes, Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing. A topological shortcut sidesteps this debate: it produces a reordering and mixing of the basis rather than an amplitude change in a fixed band, so neither low-pass nor high-pass filtering in a fixed basis is the right defensive primitive — a hypothesis our experiments test directly.
Research Program
The program has three phases, each producing a standalone finding. Phase one is a measurement study: on FAUST, SCAPE, and TOPKIDS meshes, and on ModelNet40 point clouds, we inject synthetic bridges of graded severity (controlled contact radius, single versus multiple edges) and measure eigenvalue drift, eigenfunction subspace angles, sign and order swaps, and the collapse of functional-map correspondence — quantifying, in a controlled way, the incidental-noise observations reported across the matching literature Unsupervised Learning of Robust Spectral Shape Matching, CoE: Deep Coupled Embedding for Non-Rigid Point Cloud Correspondences. Crucially, we also run GSDA++ and SAGA under this noise: the coefficient-level attacks assume the basis they perturb is the basis the victim sees, and we expect their success to degrade as topology scrambles the eigenfunctions — establishing that the two attack families are not interchangeable and motivating ours.
Phase two builds the attack on the rank-one formula: bridge-site selection from eigenvector or surrogate descriptors, mesh-edge fusion for explicit representations, and minimal point displacement (a handful of points moved sub-millimeter in normalized units) to induce $k$NN shortcuts for point clouds. We evaluate against classification (ModelNet40: PointNet, PointNet++, DGCNN, PointMLP using the official implementation pointMLP-pytorch, PointTransformer), against mesh autoencoders on CoMA in SAGA's setting, against correspondence pipelines including the diffusion-based DenoisFM, which reports a mean geodesic error of 3.5 on SHREC'19 and is known to struggle with unmodeled holes Denoising Functional Maps: Diffusion Models for Shape Correspondence, and against a unified foundation model, Uni3D, whose $k$NN-PointNet tokenizer is exactly the kind of local-connectivity mechanism a shortcut corrupts; Uni3D reports 88.2% zero-shot accuracy on ModelNet40, and no reported result found on Lacuna for Uni3D under any adversarial attack Uni3D: Exploring Unified 3D Representation at Scale. We also test cross-representation transfer in Rampini et al.'s protocol — attack designed on the mesh, delivered as a sampled point cloud — where we expect to match their transfer without their optimization cost.
Phase three turns to defense, using the attack as a probe. We test whether localized representations that the matching community adopted for topological robustness — multiscale wavelets Efficient Deformable Shape Correspondence via Multiscale Spectral Manifold Wavelets Preservation and $k$-hop neighborhoods HOPE: Shape Matching Via Aligning Different K-hop Neighbourhoods — double as architectural defenses against both our attack and GSDA++, and whether bridge-augmented retraining (the topological analog of the low-pass spectral retraining that cut GSDA++ from 100% to roughly 12.5% on PointNet Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing) buys robustness, and at what cost in clean accuracy.
Proposed Experiments
The decisive comparisons run on four benchmarks. On ModelNet40 we measure attack success rate at matched perturbation budgets — equal Chamfer and Hausdorff distance and equal count of modified points — against the published figures: GSDA++ at 100% success on five architectures, and its drop to approximately 12.5% under low-pass retraining Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing; GSDA's 100% against three architectures Exploring the Devil in Graph Spectral Domain for 3D Point Cloud Attacks. Success for us means matching or exceeding that success rate at a strictly smaller coordinate budget and lower wall-clock cost per shape. On CoMA we reproduce SAGA's setting — over 99% success forcing target reconstructions, detector accuracy of 98% on vertex attacks versus roughly 53% on SAGA SAGA: Spectral Adversarial Geometric Attack on 3D Meshes — and train the same style of detector on our attacks, predicting near-chance detection because a fused contact produces no outliers, no jitter, and no amplitude anomaly. On TOPKIDS and SHREC'19 we measure correspondence damage as mean geodesic error inflation against published robust references: HOPE's 94.9% accuracy on TOPKIDS HOPE: Shape Matching Via Aligning Different K-hop Neighbourhoods, Cao et al.'s 9.2 geodesic error on TOPKIDS Unsupervised Learning of Robust Spectral Shape Matching, and DenoisFM's 3.5 on SHREC'19 Denoising Functional Maps: Diffusion Models for Shape Correspondence. On ScanObjectNN we test whether incidental real-scan topology already produces the failures our synthetic bridges produce, connecting the attack to the sim-to-real gap documented by Uni3D and by manifold adversarial learning, which reports 73.83% accuracy on ModelNet40-to-ScanObjectNN transfer Manifold Adversarial Learning for Cross-domain 3D Shape Representation, Uni3D: Exploring Unified 3D Representation at Scale.
Baselines
The naive reference point is the do-nothing setting: each victim model's clean accuracy with no perturbation, which anchors how much damage any attack does (for Uni3D this reference is its reported 88.2% zero-shot ModelNet40 accuracy Uni3D: Exploring Unified 3D Representation at Scale). The strongest existing competitor is GSDA++, the learnable graph-spectral-filter attack that reports 100% success against PointNet, PointNet++, DGCNN, PointTransformer, and PointMLP on ModelNet40 Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing; it is the right comparator because it is the most recent, most broadly effective spectral attack on the same victims and benchmark, and because its low-frequency-constrained filter design is the closest philosophical cousin to our low-frequency-focused spectral corruption. We will run it from its published setup on the same victim checkpoints, tuning both attacks with the same budget of random restarts and holding perturbation size fixed by Chamfer and Hausdorff distance and modified-point count; beating it means equal or higher success rate at strictly lower budget, or equal success when GSDA++ fails under topological noise where ours does not. In the mesh-autoencoder setting the strongest competitor is SAGA, with over 99% success on CoMA and near-immunity to a detector that catches vertex attacks at 98% accuracy SAGA: Spectral Adversarial Geometric Attack on 3D Meshes; we reproduce its detector and its budget metrics exactly. A spatial vertex-perturbation attack serves as the naive attack baseline, expected to be caught by the detector at high accuracy, as SAGA reported for prior vertex attacks. The key ablation isolates the mechanism: a random-bridge variant that fuses a randomly chosen nearby point pair under the same coordinate budget, versus our eigenvector-guided site selection. If guided placement does not substantially outperform random placement, the claim that spectral-damage targeting matters is falsified and the phenomenon reduces to "any local corruption hurts" — an outcome the ablation is designed to detect honestly. For correspondence experiments the baselines are the robust matching systems themselves at their published TOPKIDS and SHREC'19 figures, so the quantity of interest is error inflation under attack rather than a new matching method.
Why not just…?
Why not just use GSDA++ or SAGA as the attack? These are the real competitors, and if either matches our success rate at our coordinate budget and compute cost, the attack contribution collapses. But three published facts say they will not. All coefficient-level attacks need a per-shape $O(n^3)$ eigendecomposition plus iterative optimization Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing; our attack needs one decomposition offline and a single edit. They perturb coordinates densely, spending budget on every point, while a bridge moves a handful. Most importantly, they assume the basis they perturb survives delivery — SAGA's shared spectral basis explicitly presumes consistent topology and connectivity across shapes SAGA: Spectral Adversarial Geometric Attack on 3D Meshes, an assumption the matching literature shows real scans violate Deep Functional Maps: Structured Prediction for Dense Shape Correspondence. Our phase-one experiments run both attacks under synthetic topological noise; if they hold up, that is a genuine finding against us, and the matched-budget comparison settles it either way.
Why not just defend with outlier removal and low-pass filtering? The record says these fail against structured spectral noise, and ours is more extreme: SOR and SRS are already evaded by GSDA Exploring the Devil in Graph Spectral Domain for 3D Point Cloud Attacks, and low-pass filtering actually helped SAGA by smoothing away residual high frequencies while leaving the low-frequency adversarial signal intact SAGA: Spectral Adversarial Geometric Attack on 3D Meshes. A shortcut attack changes the basis itself, not amplitudes in a fixed basis, so any fixed filter is principledmiss; still, we test the strongest version of this defense — low-pass spectral retraining, which cut GSDA++ from 100% to about 12.5% Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing — plus its topological analog, bridge-augmented retraining, rather than arguing around it.
Why not just augment training with random noise or synthetic bridges? This is the most serious alternative, because bridge augmentation is cheap and obviously worth trying; we include it as a defense baseline rather than dismissing it. Two results temper expectations. Stochastic augmentation provably underexplores structured distortions: manifold adversarial mining beat MetaSets and PointDAN on ModelNet40-to-ScanObjectNN (73.83% versus a 72.42% prior state of the art), showing worst-case structure matters, not just noise diversity Manifold Adversarial Learning for Cross-domain 3D Shape Representation. And models trained with standard augmentations remained fully vulnerable to spectral attacks — only spectral-aware retraining changed the picture Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing. Our guided-attack-versus-random-bridge ablation directly measures whether augmentation can cover the attack's support: if random bridges in training neutralize guided bridges at test time, augmentation wins and we will report that.
Why not just run the evaluation study — existing attacks under topological noise — and skip inventing a new attack? That study is our phase one and is publishable on its own, but it cannot answer the mechanism question: observing that GSDA++ degrades on noisy scans does not tell you which connectivity change causes which spectral damage, and the rank-one formula gives both a closed-form explanation and the cheapest possible instrument for probing it. The attack is the measurement tool and the defense probe, not just a headline.
Why not just use a large unified model like Uni3D and sidestep low-level geometry? Scaling does not dissolve the problem; it may inherit it. Uni3D's tokenizer is precisely a $k$NN-grouped PointNet over local patches Uni3D: Exploring Unified 3D Representation at Scale, the same connectivity a point-cloud shortcut corrupts, and unified pipelines still lose 12.9% to 39.3% of total generative error in representation conversion and reconstruction Unifi3D: A Study on 3D Representations for Generation and Reconstruction in a Common Framework. No reported result found on Lacuna for Uni3D under adversarial attack, so testing it is a genuine open question our experiments answer.
Decisive Checks
The proposal stands or falls on four numbers. First, at matched Chamfer/Hausdorff budget on ModelNet40, our success rate versus GSDA++'s published 100% Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing, with per-shape generation time reported alongside. Second, the guided-versus-random bridge ablation: without a clear gap, the spectral-targeting mechanism is dead and the honest conclusion is that local corruption, not spectral design, does the work. Third, detector accuracy on our attacks in SAGA's protocol, against its published 98%-on-vertex-attacks and roughly 53%-on-SAGA figures SAGA: Spectral Adversarial Geometric Attack on 3D Meshes; near-chance detection supports the stealth claim, high detection kills it. Fourth, cross-representation transfer without inverse-spectral optimization, compared against the 80.8% unseen-shape transfer of the universal spectral attack Universal Spectral Adversarial Attacks for Deformable Shapes. If all four land, the field gets a sobering message: the cheapest adversarial perturbation on a 3D shape is not a coordinate at all, but a connection — and every representation whose features depend on graph or spectral structure, from point-cloud classifiers to foundation tokenizers, shares the same seam.
References
- Exploring the Devil in Graph Spectral Domain for 3D Point Cloud Attacks
- Point Cloud Attacks in Graph Spectral Domain: When 3D Geometry Meets Graph Signal Processing
- SAGA: Spectral Adversarial Geometric Attack on 3D Meshes
- Universal Spectral Adversarial Attacks for Deformable Shapes
- Efficient Deformable Shape Correspondence via Multiscale Spectral Manifold Wavelets Preservation
- HOPE: Shape Matching Via Aligning Different K-hop Neighbourhoods
- Unsupervised Learning of Robust Spectral Shape Matching
- CoE: Deep Coupled Embedding for Non-Rigid Point Cloud Correspondences
- Denoising Functional Maps: Diffusion Models for Shape Correspondence
- Uni3D: Exploring Unified 3D Representation at Scale
- Manifold Adversarial Learning for Cross-domain 3D Shape Representation
- Deep Functional Maps: Structured Prediction for Dense Shape Correspondence