acceptodds
Under review as a conference paper at ICLR 2027

BadPVD: Backdoor Attacks on Point-Voxel Diffusion Models

39 views

Abstract

Diffusion models have demonstrated strong capabilities in synthesizing high fidelity and diverse data distributions, and diffusion based 3D point cloud generation methods have also achieved impressive generation performance. However, despite their strong generative ability, the security risks of these models remain largely unexplored, especially under backdoor attacks. Attackers can implant hidden backdoors into 3D point cloud diffusion models, causing the model to behave normally under benign inputs while generating attacker specified target objects once a trigger is introduced. In this paper, we investigate the vulnerability of point-voxel diffusion models to backdoor attacks and propose a noise driven backdoor injection method. By manipulating the noise generation and denoising trajectory, the model learns to associate trigger signals with attacker defined target shapes. Under trigger-conditioned initialization, the implanted backdoor can be activated while remaining inactive under the reported benign settings. Experiments under the evaluated settings demonstrate targeted generation while retaining generation capability with limited degradation on the reported metrics. These results suggest that backdoor security should be considered when deploying or reusing diffusion based 3D point cloud generation models.

Primary area:
generative models
Keywords:
backdoor attacks, point cloud generation, diffusion models, point-voxel diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iqFHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the backdoor threat against diffusion models for 3D point cloud generation. The authors design a noise-driven backdoor attack, termed BadPVD, which injects structured triggers into the initial noise distribution to steer the generation process toward attacker-specified target shapes. Specifically, the method constructs a target-guided poisoned diffusion trajectory and jointly optimizes benign and backdoor noise prediction objectives during model fine-tuning. Experiments on PVD and SPVD models demonstrate high attack success rates under trigger-conditioned initialization while largely preserving normal generation quality.

Critical strengths
  1. The authors proposed the first backdoor attack for 3D point cloud generation diffusion models. Extending the study of backdoor vulnerabilities from 3D point cloud classification and 2D diffusion models to 3D generative models is a meaningful research direction.
  2. The authors evaluate the proposed attack across multiple point cloud diffusion models (PVD, SPVD, and SPVD-C), target categories, trigger configurations, and poisoning ratios. The experimental results demonstrate high attack success rates under the evaluated settings while largely preserving benign generation quality.
Critical weaknesses
  1. Limited methodological novelty. The proposed method largely follows existing noise-triggered backdoor attacks on diffusion models, particularly BadDiffusion [1]. The target-guided poisoned diffusion trajectory closely resembles the formulation in [1], while the paper introduces few designs specifically tailored to the geometric properties of 3D point clouds or the point-voxel architecture. Therefore, the technical novelty beyond adapting existing diffusion backdoor techniques to a new generation task remains unclear.
  2. Insufficient evaluation against existing methods. The paper only compares BadPVD with a naive noise-shift baseline, without evaluating existing backdoor attacks on diffusion models. Although this may be the first work targeting 3D point cloud diffusion models, several closely related methods have been proposed for 2D diffusion models [1, 2]. In particular, the proposed poisoned diffusion trajectory closely resembles that of BadDiffusion [1]. These methods could potentially be adapted to 3D point cloud generation, or the authors should at least explain why such comparisons are not feasible. Without these comparisons, it is difficult to assess the effectiveness and technical contributions of the proposed method relative to prior work.
  3. Unrealistic threat model. The proposed attack requires the attacker to control the initial noise during inference to activate the backdoor. However, in the considered model supply-chain scenario, the attacker only modifies and redistributes the pretrained model, while the downstream user independently performs inference. It remains unclear how the attacker could manipulate the initialization noise in this setting. Moreover, if the attacker already has control over the inference pipeline, the authors should clarify why backdoor injection is necessary compared with simpler ways of manipulating the generation process. This raises concerns about the practical feasibility of the assumed threat model.
  4. Limited discussion of real-world security implications. Although the authors briefly mention potential applications such as simulation environments and digital twins, the practical use cases of pretrained 3D point cloud diffusion models remain insufficiently discussed. More importantly, it is unclear how the attacker-controlled generation of specific object categories would lead to concrete security risks in downstream applications. Unlike backdoor attacks against discriminative models, where manipulated predictions can directly affect downstream decisions, the security consequences of altering generated point clouds are less straightforward and require further justification.

[1] Sheng-Yen Chou, Pin-Yu Chen, Tsung-Yi Ho. How to Backdoor Diffusion Models? Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4015-4024

[2] Weixin Chen, Dawn Song, Bo Li. TrojDiff: Trojan Attacks on Diffusion Models With Diverse Targets. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4035-4044

Questions for authors
  1. See weaknesses above.

  2. Other minor problems:

    a. There is no section number in the paper.

    b. The organization of tables and figures on page 8 is chaotic.

    c. The related work section has been placed in appendix, which hinders readers that may not that be familiar with 3D generations and attacks.

Reviewer_GcpbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors proposed BadPVD, a noise-driven backdoor attack framework targeting point-voxel diffusion models for 3D point cloud generation. The method constructs a target-guided poisoned diffusion trajectory by introducing structured trigger signals into the noise distribution, enabling compromised models to generate attacker-specified shapes under triggered initialization while preserving benign generation capability. Experiments on ShapeNet using PVD and SPVD demonstrate high attack success rates across different target categories, trigger configurations, and poisoning ratios.

Critical strengths
  • The paper investigates backdoor vulnerabilities in diffusion-based 3D point cloud generation, which is relatively underexplored research direction.
  • The proposed method provides a clear formulation of the poisoned diffusion trajectory and explains why straightforward noise perturbation may fail.
  • The results demonstrate strong attack effectiveness under the evaluated conditions while maintaining relatively stable benign generation quality.
Critical weaknesses
  • The proposed method mainly adapts noise-triggered backdoor injection and poisoned diffusion trajectory learning to point-voxel diffusion models. Similar principles have already been explored in image diffusion backdoor methods such as TrojDiff and VillanDiffusion. Although applying these techniques to 3D point cloud generation is meaningful, the paper does not sufficiently provide direct experimental comparisons with adapted existing diffusion backdoor attacks.
  • As described in the Attack Scenario section, the attacker must redistribute a compromised model and introduce a specially constructed initial noise distribution during inference. However, in practical applications, users typically initialize diffusion models with standard Gaussian noise. The paper does not sufficiently explain how an attacker could reliably introduce the trigger during downstream deployment, which raises concerns about the practical feasibility and applicability of the proposed attack scenario.
  • Although Table 1 reports a 100% Triggered Rate (TR) across the evaluated models, the default experiments use a relatively strong trigger with a spatial offset of 2.0 and a poisoning ratio of 0.5. As shown in Table 4, reducing the trigger offset to 0.5 decreases TR to only 21.87%. These results suggest that the effectiveness of BadPVD is sensitive to trigger configuration, and its robustness under weaker or less detectable triggers remains insufficiently demonstrated.
Questions for authors

Please refer to the weaknesses.

Reviewer_2g1TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BadPVD, a noise-driven backdoor attack framework for point-voxel diffusion models used in 3D point cloud generation. The threat model is a supply-chain attack: an attacker downloads a public pretrained checkpoint, injects a backdoor via local fine-tuning, and redistributes the compromised model. BadPVD constructs a target-guided poisoned diffusion trajectory by introducing a time-dependent trigger residual, so that trigger-aware initial noise maps to an attacker-specified target point cloud while benign Gaussian initialization preserves normal generation. The authors report targeted generation with high trigger rate, low benign rate, and limited clean-generation degradation on SPVD under several trigger patterns, target classes, poisoning ratios, and sampling settings.

Critical strengths

The engineering observation that a fixed trigger added to noise accumulates into a geometry-independent offset is reasonable, and the time-dependent residual is a plausible patch. The paper also raises a timely security concern for 3D generative diffusion models. However, these strengths are local and do not yet constitute a sound or independently useful methodological contribution.

Critical weaknesses

W1. The paper claims “point-voxel diffusion models” but almost all reported experiments are on SPVD; PVD is not systematically evaluated and cross-model transferability is missing.

The title, abstract, contributions, and evaluation protocol claim “point-voxel diffusion models” and “cross-model and cross-target transferability,” but the verifiable main results, qualitative figures, trigger-offset study, poisoning-rate study, sampling study, and target-class study are all labeled SPVD. The PVD main results are either absent or unverifiable. This makes the scope of the contribution unknown. The paper must state clearly whether BadPVD targets PVD, SPVD, or both; report PVD results under the same protocol; and either provide cross-model transfer results or remove the claim. If the method only works on SPVD, the title and contributions must be narrowed accordingly.

W2. The theoretical construction is not closed.

The proposed poisoning mainly ensures that the terminal distribution of the poisoned training trajectory matches the trigger-aware initialization. The reverse sampling formula is still the standard DDPM/DDIM update, whose derivation assumes standard Gaussian noise prediction. The paper does not re-derive the reverse transition, does not prove convergence to the target, and does not analyze the accumulation error of the time-dependent offset. The authors’ own sampling results show incomplete or distorted targets at low step counts, which is direct evidence of a train–inference mismatch. The paper must either provide a rigorous derivation and conditions under which standard reverse sampling preserves the trigger, or remove the theoretical claim and reframe the method as a heuristic. W3. The naive-poisoning failure analysis is too shallow to justify the proposed design.

The static-shift baseline is compared in only one setting, and the cumulative-offset argument is sketched rather than proved. The paper does not report multiple targets, multiple source classes, multiple trigger offsets, or matched training budgets for the baseline. It therefore cannot establish that the dynamic residual is the cause of the improvement rather than a hyperparameter or training-budget effect. A proper baseline study is required.

W4. The hiddenness–effectiveness trade-off is a central unresolved problem, not a side note.

The paper’s own statistics show that effective trigger settings deviate substantially from standard Gaussian initialization. The attack succeeds when the initialization is statistically conspicuous, and becomes ineffective when the trigger is too close to the Gaussian-noise region. The paper calls this a “potential diagnostic signal” but does not discuss adaptive attacks, noise normalization or clipping, statistical detection, or the detection–evasion trade-off. For a backdoor paper, this is a core weakness.

W5. Baselines and defenses are missing.

There is no comparison with adapted 3D point-cloud backdoor methods such as PointBA, IRBA, or IBA, and no comparison with diffusion backdoor methods such as TrojDiff or VillanDiffusion adapted to 3D generation. There is no defense evaluation: input-statistics detection, noise purification, trigger inversion, fine-tuning, pruning, and checkpoint auditing are all absent. Reporting a statistical cue is not a defense analysis.

W6. The reported attack results are too uniform and lack statistical rigor.

The paper reports no random-seed variance, confidence intervals, significance tests, per-class sample sizes, or MLLM judgment variance. The distributional-separation metrics being uniformly perfect require explanation: are targets single fixed point clouds, is the target class narrow, and is the evaluation overfit? The MLLM judge uses a single GPT-4o, a single binary prompt, and a small human agreement study with unanimous labels. This is possible but suspiciously clean; the sample selection, rendering views, and judge bias need much more scrutiny. Multiple judges, multiple prompts, larger blind human evaluation, and rendering-sensitivity analysis are required.

W7. Reproducibility and presentation are below the bar.

The trigger geometry, injection-point selection rule, target point-cloud source and count, loss weighting, training epochs, optimizer, learning rate, batch size, hardware, data splits, random seeds, and SPVD checkpoint source are missing. The algorithms are inconsistent with the main-text notation. Several tables duplicate the same comparison, and figure references are misaligned. Code and data-generation scripts are only promised as supplementary material. In its current form, the paper is not independently reproducible.

Questions for authors

Q1. State explicitly whether BadPVD targets PVD, SPVD, or both.

Report PVD main results under the same protocol, including clean-generation metrics, benign rate, trigger rate, and target-class separation. If cross-model transferability was evaluated, report it; if not, remove the claim from the evaluation protocol.

Q2. Provide a formal derivation of the poisoned forward process and the reverse sampling process.

Prove or disprove that standard reverse sampling preserves the trigger-aware terminal distribution. Explain why low-step sampling produces incomplete or distorted targets, and state the conditions under which the attack is expected to remain stable.

Q3. Re-run the naive static-shift baseline under matched training budgets, multiple targets, multiple source classes, multiple trigger offsets, and multiple seeds.

Report mean/variance and failure counts. Show that the improvement comes from the dynamic residual rather than from extra samples, parameters, or training time.

Q4. Evaluate several defenses: initialization-statistics detection, noise purification, and trigger inversion or model fine-tuning.

Discuss adaptive attacks that normalize or clip the initial noise. Report the detection–evasion trade-off quantitatively.

Q5. Replace the single GPT-4o judge with multiple MLLM judges and multiple prompts, and conduct a larger blind human evaluation.

Report inter-judge agreement, rendering-view sensitivity, and per-class sample sizes.

Q6. Release code and data-generation scripts, and specify all hyperparameters, trigger construction details, target point-cloud sources, random seeds, and the SPVD/PVD checkpoint configurations.

Clarify whether geometric inputs use physical or normalized coordinates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.