acceptodds
Under review as a conference paper at ICLR 2027

MVT-Attack: Momentum Variance Tuning for Transferable Point Cloud Attacks

Abstract

The vulnerability of 3D deep learning models has garnered significant attention.However, existing point cloud adversarial attacks exhibit limited transferability due to insufficient modeling of gradient dynamics. Instantaneous gradients provide limited characterization of local optimization uncertainty, while cross-iteration gradient evolution remains insufficiently exploited, leading to unstable optimization trajectories. To address this issue, we propose MVT-Attack (Momentum Variance Tuning Attack), a novel framework that enhances attack transferability through a dual gradient regulation mechanism. Unlike previous approaches that rely on complex loss-function regularization, MVT-Attack directly regulates gradient dynamics through a two-stage optimization strategy. First, we introduce a variance-guided gradient rectification mechanism that estimates local gradient fluctuations via normal-constrained sampling and adaptively adjusts gradients to reduce source-model-specific optimization bias. Subsequently, we design a momentum optimization module that aggregates historically refined gradients to construct smoother and more consistent update directions. The proposed framework demonstrates the effectiveness of gradient dynamics regulation to improve attack transferability and achieves superior performance on the ModelNet40 and ScanObjectNN benchmarks, consistently outperforming state-of-the-art methods.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Point Cloud, Adversarial Attacks, Transferability, Momentum, Variance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sVevOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MVT-Attack, a transfer-based black-box adversarial attack for 3D point cloud classifiers. It estimates local gradient discrepancies through normal-constrained neighborhood sampling, uses the normalized correction, and aggregates corrected gradients through normalized, exponentially weighted momentum over a finite window. It further projects point-wise perturbations onto their corresponding surface normals after each update and enforces the perturbation budget. Experiments on ModelNet40 and ScanObjectNN evaluate transferability against multiple architectures, robustness under four defenses, component ablations, hyperparameter sensitivity, runtime, and a multi-scale sampling variant. The reported results show substantial transferability improvements over the evaluated baselines, although the distinct contributions of the proposed components and the evidence supporting the claimed optimization mechanism require further clarification.

Critical strengths

1. Clear motivation and a coherent optimization-level approach. The paper identifies gradient optimization for transferability as a relatively underexplored direction in 3D point cloud attacks. Rather than introducing another transfer-aware loss regularizer, MVT-Attack combines neighborhood-based gradient correction with historical gradient aggregation. Its normal-constrained sampling and point-wise normal projection are geometry-specific design choices that distinguish the method from a direct application of conventional 2D gradient attacks.

2. Broad experimental coverage. The evaluation covers two complementary datasets, two perturbation budgets, multiple source and target architectures, and four defense mechanisms. The appendix extends the source-model evaluation and includes additional target architectures, while the ablation, hyperparameter, runtime, and multi-scale sampling experiments provide useful supplementary evidence.

3. Promising reported transferability and an explicit efficiency trade-off. On ModelNet40, MVT-Attack achieves a Transferability Score of 49.0%, compared with 41.4% for CFG. For PointNet-to-PointCNN transfer under the same setting, its attack success rate is 46.0%, compared with 32.7% for CFG . The paper also reports a runtime of 0.14 minutes per sample and explicitly discusses the limited performance gain relative to the additional computational cost of multi-scale sampling.

Critical weaknesses

1. The individual contributions of the proposed components are not fully isolated. Although Table 5 provides useful ablations of variance tuning (VT), momentum (Mom), and normal-constrained sampling (NC), it omits the NC-only and Mom+NC configurations. More importantly, Algorithm 1 (line 16) introduces an additional per-iteration projection of perturbations onto surface normals, whose independent contribution is not evaluated or clearly distinguished from normal-constrained sampling. The paper also does not sufficiently clarify whether the geometric constraints and perturbation update procedures are consistent across the proposed method and baselines. These factors make it difficult to fully attribute the observed improvements to the proposed gradient optimization mechanisms.

2. The claimed optimization mechanism lacks direct empirical validation. The paper suggests that MVT-Attack guides perturbations toward flatter regions and a "Shared Adversarial Space" rather than a "Model-Specific Space" (Fig. 1), resulting in smoother optimization trajectories. However, these claims are not directly evaluated through loss-landscape analysis, cross-model gradient agreement, or optimization trajectory measurements. Furthermore, the proposed "gradient variance" in Eq. 7 is actually the difference between the neighborhood mean gradient and the current gradient, rather than a conventional statistical variance. The motivation for the normalized correction in Eq. 8 could also be better explained, particularly how its magnitude relates to gradient instability. Additional analysis would help substantiate the proposed mechanism.

3. Some empirical results would benefit from further clarification. The paper could provide more details on the number of evaluated samples to better contextualize the reported ASRs, particularly for relatively small improvements (e.g., 73.5% vs. 72.6% in Table 1). Additionally, the defense evaluation is conducted using PointNet as the only source model. On ScanObjectNN, the ASRs under SRS, SOR, and DUP-Net are consistently high across all methods (96.4–98.9%), making it difficult to distinguish their relative effectiveness.

Questions for authors

Q1: Can the authors provide direct gradient-optimization baselines and more discriminative component ablations? Please evaluate point-cloud adaptations of MI-FGSM and VMI-FGSM under matched perturbation budgets and comparable optimization settings. For the component analysis, please add an NC-only configuration and separate the effects of normal-constrained neighborhood sampling and per-iteration normal projection. Together, these experiments would clarify the novelty of MVT-Attack and identify which components account for the observed improvements.

Reviewer_951TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MVT-Attack, a transferable point cloud adversarial attack that utilizes a dynamic regulation mechanism at the gradient optimization level rather than relying on loss-function regularization, improving generalization and reducing design complexity. The attack iteratively samples neighboring point clouds with displacement ranked by alignment with local surface normals. The deviation between current and mean neighborhood gradient is computed, and a normalized version of this deviation is added back to the current gradient. Adjusted gradients are aggregated with momentum and each point’s perturbation is projected onto its surface normal under an L2 budget. The attack is evaluated on two benchmark datasets with four source models and six additional mainstream 3D paradigms. Compared to 3D-Adv, GeoA3, PF-Attack, and CFG, MVT-Attack achieves state-of-the-art transferability scores 7.6 to 13.1 points above CFG. The paper also includes results under various defenses, ablation studies, and hyperparameter sensitivity study.

Critical strengths

MVT-Attack is shown to clearly and consistently improve transferability across a wide range of models, including transformer and non-parametric targets that present the greatest challenge to generalization. The method is simple and computationally cheap (a relatively small increase in runtime compared to the state-of-the-art), which makes it a practical tool for evaluating classifier robustness.

Critical weaknesses

The method is closely related to existing transferable adversarial attacks on images. Section 2.3 describes MI-FGSM’s use of “momentum to stabilize update directions” and VMI-FGSM use of “gradient variance estimation to reduce update uncertainty and enhance transferability,” two core components of the proposed point cloud attack. However, the paper does not attempt to directly compare against these methods applied to point clouds. The ablation in Table 5 shows that variance tuning alone outperforms CFG on all targets with normal-constrained sampling and momentum adding additional gains. Including these methods as baselines would help clarify how much of the improvement comes from components specific to point clouds.

Although Section 4.1 describes the setting as an attack generated “without accessing the target model,” it appears that the hyperparameters are selected using success rates on the same target models as the evaluation. Section 4.3 states that “\mu = 0.9 demonstrates a distinct advantage on DGCNN and PCT, leading to its selection.” Tuning without victim models would better reflect the black-box setting. The claim in the abstract that MVT-Attack is “consistently outperforming state-of-the-art methods” is also somewhat overstated. CFG performs better with DGCNN as the source on ScanObjectNN at \epsilon = 0.45 (see Table 8) and under the SRS and DUP-Net defenses on ModelNet40 (as shown in Table 4).

Questions for authors

How do the transferability scores change if hyperparameters are selected using victim models that are not part of the evaluation? Is ASR computed over all samples or only over samples correctly classified by the victim prior to the attack? Under SRS, SOR, and DUP-Net on ScanObjectNN every method reaches around 96% to 99% ASR (as shown in Table 4) and Section 4 attributes this to the defended models limited feature extraction suggesting that clean misclassifications may be counted as attack success. Section 4.2 describes the perturbations as minimally perceptible but this is supported only by visualizations. Can the authors quantify this difference, perhaps by reporting Chamfer or Hausdorff distances for all methods? Were the baseline results reproduced under your exact pipeline (including the per-point normal projection) or taken from CFG?

Reviewer_gmTDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MVT-Attack, an optimization framework for improving the transferability of adversarial point clouds. It estimates a neighborhood mean-gradient deviation using candidate perturbations ranked by their alignment with local surface normals, adds a normalized correction to the current gradient, and aggregates the corrected gradients through a normalized, finite-window exponential average. The final perturbation is projected onto surface normals and constrained by a perturbation budget.

The evaluation covers ModelNet40 and ScanObjectNN, multiple source and target architectures, and several defenses. The paper also includes component ablations, hyperparameter analyses, runtime comparisons, and additional experiments in the appendix. Its main empirical contribution is improved transferability over the evaluated point-cloud attack baselines.

Critical strengths

Table 3 reports improvements over CFG of 7.6 and 13.0 percentage points on ModelNet40, and 9.4 and 13.1 percentage points on ScanObjectNN, under the two perturbation budgets. The evaluation includes several source architectures and a diverse collection of target models, with further results in the appendix. These improvements are meaningful and provide evidence that the method can be useful beyond a single source–target pair.

The method does not require training an additional generative or surrogate model. Table 5 includes ablations of variance tuning, momentum, and normal-constrained sampling. In particular, adding normal-constrained sampling to VT+Mom improves the reported DGCNN result from 62.2% to 67.6% and the PCT result from 59.4% to 64.9%. Together with the hyperparameter and runtime analyses, these experiments provide useful evidence about the contribution and practical cost of the components.

Critical weaknesses
  1. The incremental contribution over the closest optimization methods is not sufficiently established.

The estimator in Eqs. (3), (4), and (7) has the same form as the neighborhood mean-gradient deviation used by Wang and He in Enhancing the Transferability of Adversarial Attacks through Variance Tuning (CVPR 2021), which the paper already cites. Equation (9) introduces a normalized, finite-window exponential average. The potentially distinctive contributions therefore lie in the normal-ranked sampling, correction normalization and timing, and finite history window. However, the current ablations do not establish how much these modifications improve over a faithful point-cloud adaptation of VMI-FGSM with conventional momentum.

OPS, from Guo et al., Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025), is also directly relevant: it combines sampling and gradient aggregation with momentum and explicitly evaluates a point-cloud version. Its relationship to the proposed method should be addressed. Without a controlled comparison with the closest optimization approach, the substantial gains over the selected baselines do not sufficiently isolate the new methodological contribution.

  1. The point-cloud-specific sampling mechanism and geometric evaluation require more precise specification.

Normal-constrained sampling is a central domain-specific ingredient, but the candidate distribution, number of candidates before ranking, and normal-estimation and orientation procedures are not sufficiently specified. These details matter because Eq. (6) uses signed alignment with the normal. Algorithm 1 does include final normal projection and an (\ell_2) constraint; the remaining issue is to establish the exact constraint used consistently across methods and whether the NC ablation changes only neighborhood sampling or also the final projection.

The geometric-preservation claim is supported mainly by visualizations. A quantitative distortion comparison within the existing NC ablation would help establish whether the transfer gains preserve the intended geometric quality. These gaps affect the interpretation and reproducibility of the paper’s main point-cloud-specific contribution.

Questions for authors
  1. Focused validation of the incremental contribution.

Could you provide a controlled comparison with a faithful point-cloud adaptation of VMI-FGSM on a representative subset of the existing experiments, matching the loss, final geometric projection, perturbation budget, and total gradient-evaluation budget? In the same setting, replacing the finite-window aggregation with conventional momentum would help identify the benefit of the history-window design. A focused comparison would be sufficient; I am not requesting a repetition of the full benchmark suite.

  1. Clarification of sampling and geometric constraints.

What are the exact candidate-sampling distribution and candidate count, and how are normals estimated and oriented? Is the perturbation constraint applied to the entire point cloud or separately to each point, and is it identical across the compared methods? In Table 5, does disabling NC affect only neighborhood sampling, or does it also change the final normal projection? If available, quantitative geometric-distortion results for this existing ablation would materially strengthen the assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.