Do Spectral Backdoors in Point Clouds Survive Resampling?

A 3D point cloud is a finite set of points sampled from a continuous surface. The most stealthy published backdoor attack on point cloud classifiers hides its trigger in the "frequencies" of a graph built on those 1,024 specific points. A backdoor is a hidden behavior planted during training: the model classifies normally until it sees a secret pattern, the trigger, and then outputs the attacker's label. Graph frequencies are the eigenvectors of the graph Laplacian, a matrix that measures how much a function varies between neighboring points. A well-developed branch of geometry theory says that only some of these graph frequencies are real properties of the surface. The low, well-separated ones converge to the surface's own vibration modes, the eigenfunctions of the Laplace-Beltrami operator, as sampling gets denser. The rest are artifacts of which particular points happened to be drawn.

This note proposes to use that fact as a lens on 3D backdoors. The central question is simple. When a triggered object is re-scanned, meaning its surface is re-sampled at a different density or with fresh points, which part of the trigger is still there, and does the backdoor still fire? We claim that convergence theory predicts the answer. Whether a spectral trigger survives resampling is governed by how much of its energy sits in surface modes that stay stable across independent samples. That same quantity gives a principled, tuning-free purification step and a clear target for an adaptive attacker.

Core Claim

Write the trigger's displacement field $\delta$, the per-point offset the attack adds to a clean shape, in the eigenbasis of a clean shape's Laplacian as $\delta = \sum_k a_k \phi_k$. Here $\phi_k$ is the $k$-th eigenvector and $a_k$ its coefficient. Let $k^*(N)$ be the largest mode index whose eigenvector, computed on two independent $N$-point samples of the same surface, agrees after alignment above a fixed correlation threshold. This is an empirical proxy for the eigenvector stability that García Trillos et al. bound non-asymptotically; it degrades as eigengaps shrink relative to the discretization error. The quantity we track is the stable energy fraction

$$\rho = \frac{\sum_{k \le k^*} |a_k|^2}{\sum_k |a_k|^2}.$$

We make three linked claims. First, attack success after resampling is monotone in $\rho$, and triggers optimized for small point-wise $\ell_2$ distortion will have low $\rho$. The reason is that the cheapest way to move a classifier under an $\ell_2$ budget is to use high-index, sample-specific modes. Second, "reconstruct the surface, then draw fresh points" removes the unstable part by construction and needs no attack knowledge. Its cutoff comes from $k^*$ rather than from tuning. Third, an attacker forced to keep $\rho \approx 1$ must use smooth, global deformations. Those deformations are measurable as distortion of the surface's low Laplace-Beltrami spectrum, which exposes a real trade-off between stealth and resampling robustness.

Relation to Prior Work

The attack we analyze is IBAPC. It adds a learned perturbation in the graph Fourier domain of a fixed 1,024-point sample, maps it back to coordinates, and penalizes $\ell_2$ distortion. It reports distortion that is only 14.29% of the next-best attack on ModelNet40 with PointNet++. It never asks what happens at other densities, and its stealth metric ignores intrinsic geometry. Our work is a direct analysis of, and response to, this attack.

PointBA supplies spatial triggers: small rotations as low as 2.5° and inserted small shapes. It also supplies the negative result that outlier removal and jitter do not stop them. Rotations are almost entirely low-mode, with $\rho$ near 1. So PointBA serves as our control, a trigger we predict survives resampling.

The theory we reuse comes from a connected set of results. Belkin et al. split convergence into a sampling stage and a bandwidth stage. Hein et al. prove convergence under non-uniform sampling, which matters because triggered shapes are sampled from deformed surfaces. García Trillos et al. give the spectral rates. We use these only to motivate $k^*$; we do not claim new theorems.

Wang et al. show that graph filters converge to manifold filters, and A Manifold Perspective on the Statistical Generalization of GNNs ties generalization to filter smoothness. Both explain why DGCNN, which builds nearest-neighbor graphs internally, should behave consistently across densities on clean data. Both study only clean data; we ask whether the backdoor inherits that consistency.

The defense literature constrains our design. Detecting Backdoor Attacks Against Point Cloud Classifiers (RED) searches for localized clusters of trigger points, so global spectral triggers fall outside its search. Spectral Signatures looks at the spectrum of feature covariance rather than of the input geometry, and needs a fully trained, already compromised model. The edge-dropping defense RIGBD is designed for relational graphs. On dense geometric graphs, dropping edges only perturbs a discretization that still converges to the same operator.

Research Program

Measurement. For each attack, compute $k^*(N)$ per shape by drawing independent samples from the original ModelNet and ShapeNetPart meshes. Then project each trigger onto the stable and unstable subspaces to obtain $\rho$.

Resolution transfer. Take triggered surfaces and re-draw the points, which requires a surface. For IBAPC, we fit a surface to the triggered 1,024 points by moving-least-squares reconstruction. We then measure attack success rate (ASR) and clean accuracy at $N' \in \{512, 1024, 2048, 4096\}$, with the model fixed at its training resolution. We also test with fresh points at $N' = 1024$, which isolates the effect of the point identity changing from the effect of density changing.

Purification. The deployable defense is reconstruct-and-resample. It fits a surface with bandwidth matched to the scale of $k^*$ and draws fresh points. As a variant, the bandwidth is chosen by the InfoGain criterion from Manifold filter-combine networks.

Adaptive attack. We re-optimize IBAPC's trigger restricted to the stable subspace. We then report its ASR, $\ell_2$ distortion, and a new intrinsic stealth metric: relative change in the first $k^*$ Laplace-Beltrami eigenvalues. This probes whether resampling-robust triggers can remain invisible.

Sharp CAD edges and boundaries violate the smooth-manifold assumptions behind the theory. We treat this as an empirical question and report $k^*$ separately for shapes with and without large boundary regions.

Baselines

The naive reference is no defense on test inputs at the training density. This is where IBAPC and PointBA report their numbers: PointPBA exceeds 95% ASR at a 5% poisoning rate with under 2% clean accuracy drop (PointBA). The IBAPC paper describes its ASR only as competitive, a few points below the best attacks; no exact number for it is reported in the file. We therefore re-run IBAPC ourselves on PointNet, PointNet++, and DGCNN with its published objective, and train every victim model with identical data and schedules.

The strongest existing detector for point clouds is RED (Detecting Backdoor Attacks Against Point Cloud Classifiers). It reports correct detection of 6 of 7 attacks on ModelNet40 with zero false positives at a significance threshold of 0.05, using 10 samples per class. We run it with the same 10 samples per class. Our defense beats it if it reduces IBAPC's ASR to near clean-model misclassification levels on attacks RED misses, at a comparable clean accuracy cost.

Spectral Signatures reports poisoned misclassification falling to within 1% of a clean model on CIFAR-10. No reported result was found on Lacuna for it on point clouds, so we apply it to penultimate features of each point network with the paper's removal fraction. Statistical outlier removal and random point subsampling are included because PointBA reports them failing; they check that our pipeline is not simply stronger than theirs.

The ablation that isolates the key ingredient is a matched low-pass filter. It projects the same 1,024 points onto their own first $k^*$ graph eigenvectors, with no reconstruction and no new points, at a cutoff swept over a grid tuned on held-out clean data. Reconstruct-and-resample must beat this filter, or match it without tuning, for the resampling step to count as a contribution. All defenses are given the same clean validation budget for any hyperparameter search.

Why not just…?

Why not just apply a spectral low-pass filter to the point cloud? This is the real threat. Low-pass filtering removes high-index modes, which is most of what we predict matters. Two things separate it from our method. Its cutoff is a free parameter an attacker can anticipate, whereas ours is set per shape by $k^*$. It also keeps the attacker's exact point positions, while resampling discards them. The matched ablation above settles the question. If tuned low-pass filtering equals reconstruct-and-resample, we will report that the defensive value of our approach reduces to choosing the cutoff from convergence theory. The measurement and resolution-transfer findings still stand on their own.

Why not just subsample or jitter points? PointBA reports that outlier removal and spatial augmentations leave its attacks at up to 100% ASR, and IBAPC reports evading augmentation-based defenses. Subsampling keeps a subset of the attacker's displaced points and therefore preserves much of the trigger, including its high-frequency part. Our resolution-transfer experiment reports subsampling and fresh resampling side by side to show this directly.

Why not just use Spectral Signatures or RED? RED searches for a localized trigger cluster by construction, and IBAPC's perturbation is distributed over the whole shape. Spectral Signatures needs a trained, compromised model and assumes the trigger dominates one feature direction. Neither prevents triggered inputs from reaching the model at test time. We run both, so any case where they already suffice will appear in the results.

Proposed Experiments and Decisive Checks

The experiments use ModelNet40, ModelNet10, and ShapeNetPart, with PointNet, PointNet++, and DGCNN as victims. The attacks are IBAPC, PointPBA rotation and interaction triggers, the PointBA clean-label variant, and our stable-subspace adaptive attack. Metrics are ASR, clean accuracy, $\ell_2$ distortion, low-spectrum eigenvalue distortion, and $\rho$.

The first decisive check tests whether ASR under fresh resampling tracks $\rho$ across attacks and shapes. If IBAPC's ASR collapses under resampling while rotation triggers persist, and a rank correlation with $\rho$ holds within attacks, the core claim is supported. If IBAPC survives resampling despite low $\rho$, the claim is false. That outcome would mean the network responds to something other than the spectral content we measured.

The second check tests whether reconstruct-and-resample beats the tuned matched low-pass filter at equal clean accuracy.

The third check tests whether the adaptive stable-subspace attack can recover IBAPC-level ASR only by incurring measurably larger low-spectrum distortion. If it recovers that ASR without such distortion, our defense can be evaded, and we will report the resolution-transfer analysis as the main result.