acceptodds
Withdrawn from ICLR 2027

Rethinking High-Frequency Forensic Signals for Poison-Splat Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Poison-Splat exploits adaptive densification in 3D Gaussian Splatting (3DGS) to increase reconstruction cost through perturbed training images. Recent defenses use high-frequency patterns as forensic evidence of poisoning and report strong performance on standard benchmarks. We question whether these signals provide a reliable fingerprint of Poison-Splat when legitimate scene content itself contains texture. To investigate, we introduce TextureBench, a benchmark of ordinary, naturally occurring scenes with legitimate fine-scale texture, paired with valid clean 3DGS reconstructions and matched poisoned views. Using this benchmark, we analyze three aspects of defense reliability: detection, preservation of clean content during purification, and their interaction in composed pipelines. The RemedyGS detector's AUROC drops from 0.980 on existing benchmarks to 0.617 on TextureBench, while several high-frequency statistics reverse their clean–poisoned ordering. Purification also causes larger reconstruction losses on clean textured scenes, and false-positive detections propagate this damage through the composed defense. Finally, under the same task-specific training supervision, a detector using frozen DINOv2 features improves TextureBench AUROC to 0.904 and reduces downstream damage with the purifier held fixed, although substantial errors remain. Our results prompt re-thinking of using high frequency patterns for Poison-Splat detection. We include the code and benchmark in the submission.

Primary area:
datasets and benchmarks
Keywords:
3D Gaussian Splatting, Poison-Splat Defenses, Natural Texture

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_co4vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the reliability of high-frequency forensic signals for Poison-Splat detection in 3D Gaussian Splatting. The authors introduce TextureBench, a benchmark containing 28 real-world texture-rich scenes, to examine whether existing defenses can distinguish natural high-frequency textures from poisoning artifacts. Experiments show that existing detectors suffer significant performance degradation on TextureBench and that unnecessary purification can damage clean reconstruction quality. The paper further demonstrates that replacing the original detector with a DINOv2-based classifier improves detection performance and reduces reconstruction damage while maintaining attack mitigation.

Critical strengths
  1. The paper introduces TextureBench, a benchmark comprising 28 real-world texture-rich scenes, providing a new testbed for evaluating the generalization of Poison-Splat detectors under complex natural textures.

  2. The paper reveals substantial performance degradation of existing detectors on texture-rich scenes. In particular, the AUROC of RemedyGS drops from 0.980 to 0.617, highlighting generalization limitations under challenging real-world conditions.

  3. Beyond detection accuracy, the paper investigates how false positives trigger unnecessary purification and degrade 3DGS reconstruction quality. It also demonstrates that DINOv2-based detection can reduce reconstruction damage on clean scenes.

Critical weaknesses
  1. Finding 1 is not sufficiently supported. A reversal in average high-frequency energy does not imply that natural textures and poisoning perturbations lack discriminative differences in their high-frequency distributions. RemedyGS itself employs a learning-based CNN detector rather than simple frequency statistics. The effective combination of a DINOv2 detector and the RemedyGS purifier also suggests that the original defense pipeline remains useful, with the performance bottleneck potentially lying in feature representation rather than high-frequency forensic information itself.

  2. Finding 2 offers limited novelty. The potential degradation of natural textures and reconstruction quality caused by purification has already been discussed in RemedyGS, and this concern is a key motivation for its selective purification mechanism.

  3. Finding 3 lacks a convincing explanation of the underlying mechanism. DINOv2 may also exploit complex high-frequency forensic features. Its advantage over the CNN-based RemedyGS detector could primarily result from large-scale pretraining and stronger feature representations, rather than reduced reliance on high-frequency information.

Questions for authors

The three findings require a deeper analysis of the underlying mechanisms, as some interpretations appear inconsistent with the experimental results. In particular, the effective combination of DINOv2 detection and RemedyGS purification suggests that the original defense pipeline remains useful. The performance improvement may stem from stronger pretrained representations rather than reduced reliance on high-frequency information. The paper does not clearly distinguish the limitations of simple high-frequency statistics from those of high-frequency forensic signals themselves. Without this distinction, the work appears to be primarily a new benchmark and a backbone replacement, repackaged into a broader narrative about the limitations of high-frequency forensic signals.

Reviewer_7thAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper re-examines the reliability of high-frequency features for Poison-Splat attack detection and introduces TextureBench, a benchmark consisting of 28 naturally textured scenes. By evaluating attack detection, image purification, and their combined defense pipelines, the authors demonstrate that existing high-frequency-based defenses may misclassify legitimate textures as attacks, leading to degraded 3DGS reconstruction quality. Furthermore, the authors investigate detectors based on pretrained MAE and DINOv2 features, demonstrating the potential of general-purpose visual representations to improve detection robustness.

Critical strengths
  1. Meaningful research problem. The paper identifies a potential limitation of existing Poison-Splat defenses that rely on high-frequency features, providing a valuable perspective on evaluating defense reliability.

  2. Comprehensive experimental evaluation. The authors systematically investigate attack detection, purification of clean images, and complete defense pipelines, effectively demonstrating how false positives affect downstream reconstruction quality.

  3. Insightful experimental findings. TextureBench reveals substantial performance degradation in existing defense methods, while the results obtained with DINOv2 provide useful insights for future research.

Critical weaknesses
  1. Limited methodological novelty. The paper primarily focuses on constructing a challenging benchmark and evaluating existing pretrained visual models, without introducing a substantially new detection or defense mechanism.

  2. Limited benchmark scale and diversity. TextureBench contains only 28 carefully selected static textured scenes. Its representativeness remains uncertain, and the findings may not generalize to more complex real-world environments.

  3. Insufficient methodological analysis. Although DINOv2 significantly improves detection performance, the paper does not adequately explain the underlying reasons for these improvements or systematically investigate the contributions of different visual representations and detection strategies.

  4. Limited attack evaluation. The experiments primarily focus on Poison-Splat, without systematically evaluating different perturbation strengths or other related attacks, which limits the generality of the conclusions.

Questions for authors
  1. Benchmark representativeness: Given that TextureBench contains only 28 scenes, how do the authors justify its representativeness of naturally occurring textures in real-world 3DGS applications?

  2. Effectiveness of DINOv2: What specific characteristics of DINOv2 features contribute to its superior performance over RemedyGS and MAE? Do the authors have additional analyses that could explain these differences?

  3. Robustness across attack settings: Do the observed detection failures and the advantages of DINOv2 remain consistent across different Poison-Splat perturbation strengths?

Reviewer_bug6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper re-examines the Poison-Splat defense method based on high-frequency features, proposing the TextureBench containing 28 real-world texture scenes. It finds that natural textures easily lead to false alarms in existing detectors and impair the quality of normal reconstruction during the cleanup process. Experiments show that the DINOv2-based detector can improve generalization performance and reduce the losses caused by false alarms.

Critical strengths
  1. This paper is easy to read.
  2. The experiments are comprehensive and convincing.
  3. The findings and observations in this paper are properly justified.
Critical weaknesses
  1. The major weakness that support rejection is the limiations of novelty and contribution. The major findings and observations mostly have been discussed in the area of adversarial machine learning. Specifically, features correlated with attacks are not necessarily attack-specific, as discussed by Carlini and Wagner (2017); the trade-off between robustness and clean-input performance has been explored by Tsipras et al. (2019); and the ability of naturally occurring hard examples to expose failures overlooked by standard benchmarks has been demonstrated by Hendrycks et al. (2021). Although the paper provides empirical evidence of these issues in the context of Poison-Splat defenses, it does not sufficiently clarify what new, Poison-Splat-specific failure mechanisms or insights are revealed beyond confirming known limitations in a different setting.

  2. While the DINOv2-based detector achieves promising improvements on TextureBench, the paper provides limited insight into why it performs better. It remains unclear whether the improvement stems from better distinguishing legitimate textures from attack artifacts or simply from stronger pretrained representations. Without further analysis of the underlying mechanisms, this result offers limited guidance for developing more effective Poison-Splat defenses.

  3. Finding 3 appears somewhat intuitive: reducing false positives naturally avoids unnecessary purification and thus better preserves clean-input quality. Therefore, I think this finding offers limited insight.

[1] Carlini, N., & Wagner, D. (2017). Adversarial Examples Are Not Easily Detected: Bypassing Ten Detection Methods. Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security.

[2] Tsipras, D., Santurkar, S., Engstrom, L., Turner, A., & Madry, A. (2019). Robustness May Be at Odds with Accuracy. International Conference on Learning Representations (ICLR).

[3] Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., & Song, D. (2021). Natural Adversarial Examples. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 15262–15271.

Questions for authors

If my assessment about the contribution and novelty mentioned before is incorrect, what are the actual novelty and contribution of this work?

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

Related papers

Discussion (0)

Sign in to comment.