acceptodds
Under review as a conference paper at ICLR 2027

SplatLeak: Tracing and Reducing Membership Leakage in 3D Gaussian Splatting

Abstract

A 3D Gaussian Splatting (3DGS) asset can be shared without its source photographs. Can its renders reveal which photographs trained it, when neighboring views can reconstruct images that never supervised optimization? We isolate the fidelity gained from a photograph's own supervision using shared-initialization and matched-schedule interventions, even with three synchronized neighboring views already supervising. We audit membership by comparing target errors with those of same-scene references trained without the candidate pool. On eight new test scenes, a photometric log-ratio improves mean normalized partial AUC over 0–1% FPR by 0.1567 relative to PSNR-GAP. We also audit candidates without supplied poses using registered background images and cameras with known candidate intrinsics. Reference-fitted alignment extends this audit across the tested numerical color changes. To reduce source association before release, we distill students from teacher renders. Ordinary distillation retains teacher-source association even without direct supervision from protected photographs. We instead supervise each protected photograph's registered camera using teachers that excluded it. Across four scenes, this lowers the mean per-asset maximum AUC over nine attack configurations from 0.9457 to 0.7364, at a mean held-out PSNR cost of 0.241 dB relative to membership-agnostic distillation. The student retains higher held-out PSNR and SSIM than training on auxiliary photographs alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
membership inference, 3D Gaussian Splatting, neural rendering, privacy auditing, source-exclusion distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_2uEoOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In general, this paper introduces a splat eak, which can identify whether one image is used to construct the 3DGS. The author also discusses how to reduce this potential private release problem. Overall, I think the introduced idea is interesting and could be useful in some scenarios when one image owner wants to verify whether his/her image is used by others to reconstruct 3D assets.

Critical strengths

(-) The idea itself is interesting, and I think this is the first study of such a membership inference problem in image-3DGS. I think some readers could find this paper useful and have some following works.

(-) Basically, it designs a method to find a candidate image's camera position. A naive solution to find the position could directly inverse render and continue to optimize the camera position to find the best similar scores, but it fails as it is hard to optimize. The introduced methods seem useful: 2D-3D matching and PnP estimation. Calculating photometric log-ratio just passed the estimation errors.

(+) Also studied how to avoid such a private release problem.

Critical weaknesses

(-) The proposed auditing framework relies on training multiple same-scene background reference models (default K=8K=8), while the defense pipeline similarly requires pre-training 4 teacher models and generating synthetic supervision renders. In large-scale scenes or compute-constrained environments, this cumulative training overhead may pose a practical bottleneck. Under severe computational or time constraints, if the reference budget is compressed to fewer models (e.g., K=2K=2 or K=4K=4), what is the degradation trajectory regarding detection recall and false positive rate (FPR)? Furthermore, is there a lower-cost single-model calibration alternative viable for lightweight 3DGS deployments?

(-) When candidate poses are unknown, the localization pipeline relies on a registered set of public background photographs to construct a sparse 3D point cloud map via feature matching and PnP. If an adversary severely crops the background regions or if the auditor cannot obtain overlapping background imagery from the same scene, this feature-matching-based relocalization may fail. In cases with extremely sparse texture (where extracting sufficient inlier matches is challenging) or when registered background images are unavailable, is there a feasible strategy to directly invert the camera pose against the 3DGS radiance field without relying on explicit geometric mapping?

(-) The gray hline in Figure 1 could be darker, as the current version does not clearly divide into different parts enough.

Overall I think the idea is novel, and the experiment is compressive generally. I would keep my positive rating unless I see other serious weaknesses from other reviewers.

Questions for authors

Please see above.

Reviewer_TqS7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies training-image membership inference, which determines whether a photograph was used for training based on the rendering results of a 3DGS model. The authors use shared-initialization and matched-schedule interventions to verify that an image’s own photometric supervision improves its reconstruction, and then infer membership using a photometric log-ratio to compare reconstruction errors between the target model and same-scene reference models trained without the candidate pool. In addition, the authors use source-aware teacher routing, where teachers that excluded the corresponding source image provide supervision, to train a student with lower source association and reduce membership leakage from the trained 3DGS. They evaluate the approach across multiple datasets and experimental settings.

Critical strengths
  • The experiments are comprehensive. Through various settings and controlled experiments, the paper systematically studies training-image membership leakage in 3DGS and ways to solve it.
Critical weaknesses
  • The motivation seems unclear.
  • The threat model has limited applicability.
Questions for authors
  • The motivation seems unclear. The authors define the task of inferring individual-image membership from 3DGS, but do not explain what information this inference reveals or what threats it poses. To my knowledge, this specific task has no prior work, so the authors should explain its motivation more fully, including attack scenarios to show why the problem needs to be solved.
  • The threat model may has limited applicability. Someone who obtains a 3DGS model may not have access to candidate intrinsics or background images. A setting worth considering is membership inference with only the candidate image and the 3DGS model, without this additional information.
  • The comparability of experiments are unclear. Different experiments use different scenes without a consistent standard, making it difficult to determine whether differences in results come from the method or the experimental settings. Comparisons on a consistent set of datasets would be preferable.
Reviewer_UGtnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies image‑level membership inference auditing for published 3D Gaussian Splatting assets, where the auditor can only obtain rendered outputs without direct access to model weights. The authors design well‑controlled causal intervention experiments to verify that a photograph used for photometric optimization creates measurable fidelity improvement at its own camera viewpoint. Built upon existing log‑ratio membership‑inference frameworks, they propose the log‑photo readout, together with pose‑estimation and gamma+RGB photometric alignment to handle real‑world color perturbations. Further, they present an OUT‑routing distillation defence to mitigate membership leakage. Extensive experiments across main text and appendix are provided for evaluation.

Critical strengths
  1. Novel practical research problem & high‑quality causal intervention:This paper brings image‑level membership‑inference auditing to 3DGS, an under‑explored privacy problem for widely‑distributed neural‑rendering assets. Section‑4 controlled intervention studies cleanly isolate the fidelity trace introduced by photometric supervision.

  2. Complete attack‑and‑defend workflow:The paper builds a full pipeline including auditing readout, pose‑estimation, perturbation robustness, and distillation‑based mitigation, rather than only evaluating attacks.

Critical weaknesses
  1. Empirical scope of the core claim is uncharacterised, and the main evaluation conflates two leakage channels. The paper's central claim — that a photometric-supervision member leaves a measurable fidelity gain at its own camera view — rests on causal-intervention observations, yet no boundary analysis is given for when it may fail (e.g., low-gradient, redundant, or noisy views; early-stopped training), i.e., cases where an image participates but influences the model little. More importantly, the formal membership definition in Section 3.1 concerns photometric optimisation only, while in the Section-5 main benchmark member images enter both SfM initialisation and photometric optimisation; the reported leakage therefore mixes initialisation traces with photometric-supervision traces, undermining the construct validity of the headline results.

  2. OUT‑routing defence lacks scalability and formal analysis of multi‑image indirect cross‑sample leakage. A key unaddressed failure arises when protecting multiple images e.g. (P1, P2) at once. Teachers generating supervision for (P1) exclude (P1) but are trained on other protected images like (P2) , and vice‑versa. As Section 4 confirms, 3DGS can reconstruct viewpoints from neighbouring views, so residual membership traces remain even after excluding the target image. This mechanism is seen in experiments but receives no formal analysis. Additionally, with 4 teachers and 2 exclusions per image we only get 6 unique masks. Protecting more than six images forces mask reuse and amplifies indirect leakage. The defence is only tested across four scenes, leaving large‑scale performance under‑validated.

Questions for authors

Q1. Please provide ablation, grid-search, and sensitivity analysis for the hyper-parameters of the log-photo score, and a statistical or mechanistic explanation of why the log transform improves discrimination. As presented, the design appears heuristic, and its incremental contribution over existing log-ratio MIA readouts is hard to assess without these. Q2. In the alignment module, fitting applies the nonlinear gamma transform to the pixel-average of the 8 reference renders, whereas scoring applies gamma per reference image and then averages the loss. Are these two operations equivalent? Q3. The decision threshold is calibrated on a specific negative pool; how does it generalise across scenes, datasets, and capture conditions? Please analyse threshold transportability and consider a distribution-free or conformal calibration scheme.

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

SplatLeak studies whether renders from a released 3D Gaussian Splatting model reveal which photographs supervised its training. A photograph can be reconstructed accurately even if it was held out, because nearby views may cover the same surfaces. The authors isolate the effect of a photograph’s own supervision by fixing initialization and the training schedule and switching its photometric gradient on or off. They then audit membership by comparing the target model’s reconstruction error with that of same-scene reference models trained without the candidate photographs. A photometric log-ratio improves average low-FPR ranking over a PSNR-based baseline on held-out scenes.

The paper also examines whether a distilled student reveals which photographs trained its teachers. The authors find that this association can persist even when the student never sees the original photographs. To reduce it, they train the student at each protected camera using renders from teachers that excluded the corresponding photograph. Across the evaluated attacks, this reduces source association relative to ordinary distillation, with a modest average loss in held-out rendering quality and residual leakage on some assets.

Critical strengths
  1. The controlled experiments isolate the contribution of a photograph’s own supervision. The authors hold initialization, camera order, and the optimization schedule fixed, then switch the candidate photograph’s photometric gradient on or off. Even with three synchronized neighboring views already supervising both models, enabling the candidate’s gradient improves its own-view PSNR in all eight tested units. This addresses a central ambiguity in same-scene membership inference: neighboring photographs can make a candidate easy to reconstruct even when it never supervised training.

  2. The comparisons identify what each proposed method contributes. The attack compares log-photo and PSNR-GAP on the same renders and references, so the ranking difference can be attributed to the choice of error measure. The defense comparison fixes the teacher bank, student initialization, camera roster, and training schedule while changing which teachers supervise each protected camera. Fitting attacks separately for each student recipe provides evidence that teacher selection reduces source association across several ways of measuring it.

Critical weaknesses

1.The precision of the reported low-FPR gain is limited by the number of nonmembers per scene. At the 1% endpoint, each scene’s result depends on its highest-scoring nonmember, and omitting one image can reverse the ordering of log-photo and PSNR-GAP in some scenes. Threshold transfer poses a separate concern. Changing the calibration scenes materially changes the number of false positives on the same test set, limiting confidence in the operating point on a new scene.

  1. Log-photo retains the relative-error structure of PSNR-GAP, and OUT adapts a teacher-exclusion principle explored in KCD and SELENA. The comparison with ordinary distillation isolates the benefit of routing within the same teacher bank. Its practical advantage over cheaper alternatives remains unclear. The early-stopping baseline uses a different audit protocol, so the reported results do not establish how much additional protection routing provides at comparable rendering quality.
Questions for authors

1.How precisely can the current data estimate the low-FPR ranking advantage, given its sensitivity to the highest-scoring nonmember in each scene? Separately, given the variation across calibration-scene subsets, what false-positive rate should an auditor expect when applying the frozen threshold to a new scene?

  1. Suppose an attacker knows the exclusion rule and can train shadow models using the same recipe, while target access is limited to renders from the released student. Could the pattern of residuals at a protected camera and nearby cameras reveal information specific to teacher routing beyond what log-photo and the existing local-response probe capture?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.