- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes Anisotropy-Gated Densification (AGD) as a defense against Poison-splat, which is a computation-cost attack that perturbs training images to inflate 3DGS densification. AGD computes a tile-level spectral anisotropy score from an amplitude-weighted orientation tensor over mid-band FFT coefficients. Experiments are done on NeRF-Synthetic, Mip-NeRF360 and Tanks&Temples. Under poisoning, AGD reduces peak Gaussian count by 17–28%, peak memory by 14–51% and training time by 15–23% relative to undefended 3DGS, at a cost of ≤0.35 dB PSNR. Clean-scene quality is essentially unchanged. The paper also combines AGD with three existing defenses.
Critical strengths
-
The research problem is interesting and important: Compute-cost attacks matter for 3DGS training services. Regulating the densification decision directly is a reasonable complement to input purification.
-
Near-zero clean-scene cost. Clean quality is nearly unchanged with the proposed defense.
-
Lightweight and plug-in. Spectral maps are precomputed and the whole pipeline does not require learning. The method can be integrated into four 3DGS variants.
Critical weaknesses
-
The core mechanism of the proposed method lacks interpretation. STI aggregates tile-level anisotropy with a median over tiles (Eq. 5) and then over images (Eq. 6). Why use median but not other statistics? What is the advantage? If the design is known by the attacker, would it be easy to evade (an adaptive attack)? These methodology choice should be discussed and analysed more in-depth.
-
The effectiveness of AGD is not strongly supported by the results. From Table 1 and Table 2, we can see that AGD alone is weaker than existing defenses (comparing result of AGD in Table 1 and result of all three other defenses in Table 2), e.g., #Gaussians (M) of AGD alone is 0.561 while the #Gaussians (M) of image smoothing, remedy GS and DefenseSplat are all under 0.3. Additionally, adding AGD to RemedyGS increases #G, memory and time under poisoning (see Table 2, the row of RemedyGS)
-
Regarding removing needle-like Gaussians: how to define and distinguish needle-like, plate-like and sphere Gaussians? Are there only 3 categories in total? Removing needle-like Gaussians exhibit similar (or even worse) degradation in PSNR compared with random removal, it seems that such needle-like Gaussians removal can be further improved by finer-grained detection or specification. If the removal of Gaussians will cause a degradation less than random removal, the design will make sense more. SR's memory saving effect also seem to only appear in Nerf-Synthetic, but not significant on MIP and TT (Tables 10-11).
Questions for authors
-
Choice of median. What is the rationale for using the median at both levels: over tiles (Eq. 5) and over images (Eq. 6)? Please consider comparing it against alternatives such as the mean, a trimmed mean, a high percentile (e.g., 90th), or the max.
-
Adaptive attacks. How does AGD perform against an attacker who knows the design? In particular: (a) poisoning fewer than half of the training views, so the image-level median is unchanged; (b) confining the perturbation to fewer than half of the tiles in each image, so the tile-level median is unchanged; (c) adding a differentiable penalty to the Poison-splat objective that keeps below .
-
Why does adding AGD to RemedyGS increase cost under poisoning?
-
How exactly are needle-like, plate-like and spherical Gaussians defined, e.g., which thresholds on the ratios between scale axes? Are these three categories exhaustive and mutually exclusive, or do some Gaussians fall outside all three? What fraction of Gaussians falls into each category in clean vs. poisoned scenes?