- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes a defense ValidGS against the StealthAttack (which injects fake Gaussians to render a target appearance in the poisoned view). The idea is to utilize multi-view consistency: find image regions/patches with low neighboring consistency support, and replace this area with neighboring colors, and then map the suspicious and trusted image patches back to Gaussians. Trusted Gaussians are protected and suspicious Gaussians are deleted from time to time.
Critical strengths
The paper addresses view-dependent poisoning of 3DGS, a recent and under-explored security risk. The design of methodology is sensible and the evaluated settings are broad.
Critical weaknesses
-
Several rendering-quality numbers are implausible. On Mip-NeRF360, ValidGS trained on clean data reaches 26.86 dB PSNR (Table 14), whereas the same method trained on poisoned data reaches 30.31 dB V-TEST PSNR on innocent views (Table 2). It is unclear how a defense can render innocent views 3.5 dB better when trained on poisoned data than on clean data. The same pattern appears elsewhere. The undefended Poisoned 3DGS (27.76 dB) already exceeds clean Vanilla 3DGS (27.21 dB). On Free, ValidGS raises V-TEST PSNR from 25.25 to 30.47 dB, even though the attack objective (Eq. 3) explicitly keeps innocent views close to the clean model, so a defense should at best recover clean-model quality. A V-TEST PSNR of 30.31 dB is also well above what state-of-the-art methods report on Mip-NeRF360 under clean training (around 28.5 dB, e.g., Zip-NeRF). Please clarify exactly how V-TEST is computed.
-
The compared baselines are not designed for defending such attack: MaskGaussian targets model compaction; Deng et al. (2024) targets efficient density control; RemedyGS defends against computation-cost attacks (Poison-splat). Their failure in this setting is expected. Meanwhile, robust reconstruction methods for transient distractors, such as RobustNeRF (Sabour et al., 2023), SpotLessSplats, and WildGaussians seem to be more proper as baseline, since they rest on the same principle as Stage 1 of ValidGS: identifying image content that is not supported by the other views and down-weighting it. An object that appears in only one training image is, from their perspective, exactly a transient distractor. It will be helpful for the readers to understand the difficulty of this problem by reporting their performance.
-
The reliability of the clean reference in the poisoned region is not analyzed. Stage 1 builds the clean reference by unprojecting pixels with the current rendered depth and reprojecting them into neighboring views. In the injected region, however, the rendered depth is largely determined by the poisoned Gaussians. StealthAttack places these at low-density locations along the target ray, in front of the true surface (Appendix A.2). Reprojecting with this depth samples the neighboring images at parallax-shifted locations, so the fused color comes from a different part of the scene rather than from the occluded background the reference is meant to recover. The reliability weight w(x) may down-weight some of these pixels, and the depth may become correct once the illusion is suppressed. However, references archived early in training could lock in incorrect colors. The paper does not discuss this.
Questions for authors
-
How exactly is V-TEST computed? How does ValidGS trained on poisoned data outperform the same method trained on clean data by about 3.5 dB on Mip-NeRF360?
-
How do distractor-robust methods such as SpotLessSplats and WildGaussians perform under the same attack setting and protocol as Tables 1 and 2? If they fail, why?
-
Within the injected region, how accurate is the fused reference I_ref compared with the clean target image, especially early in training, when the poisoned Gaussians still dominate the rendered depth and references may already be archived?