- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Previous:
Safe4D studies causal video perturbations against unauthorized 4D reconstruction. The evidence mainly supports degradation of predicted depth and geometry, with transfer to some unseen models. Robustness to compression and purification, runtime, and effects on final rendering remain limitations.
Update:
This paper propose to use the protective effect of causal video perturbations on feed-forward reconstruction models, with evidence primarily drawn from the observed increase in depth and geometric errors. Experiments reveal that the proposed protective effect can generalize to some OOD models. But, there are still some limitations in video compression and rendering quality, etc.
Critical strengths
Previous:
The problem formulation and causal setting are valuable. The added transfer experiments strengthen the evidence: depth error increases on 61/62 sequences for VGGT and 62/62 for STream3R without receiver-specific tuning. The temporal-filtering results also support robustness to the tested filters.
Update:
It is a pioneering work on protecting video from 4D reconstruction, and extensive works about feed-forward 4D reconstruction models are compared in the appnedix. This experiement reinforces the evidence for cross-model transferability.
Critical weaknesses
Previous:
Transfer remains model-dependent, with inconsistent results on CUT3R. H.264 compression substantially weakens protection, and DiffPure nearly removes the effect on the single tested clip, although resizing alone also reduces it. The viewing test does not support an unconditional imperceptibility claim. Runtime remains high, and the added all-camera 4DGS experiment does not show consistent rendering degradation. Geometry-level disruption therefore does not yet establish reliable end-to-end protection.
Update:
When using direct input, depth error increased, with the average change being slightly negative. Consequently, transferability of the proposed method is highly dependent on the selected model.The receiver and agent models share the same encoder, but CUT3R does not. Therefore, current experiments do not treat the encoder as a variable.
Questions for authors
- Can the author report the full distribution of transfer results (not just the mean) and analyze why transfer fails on some sequences?
- Is there a path to reduce the per-frame cost by 2–3 orders of magnitude (e.g., an amortized perturbation generator)?
- Could you add a small human annotation to prove the experiental insights?
10 comments
Response to Reviewer 9Df6 (Part 2/5)
Q2: How much protection survives temporal smoothing or DiffPure? Does the +1.94 dB result imply that the attacker can bypass the protection?
Response 2: Thank you for the careful reading, and we apologize for the confusion about the result. We answer the two parts in turn.
(1) Temporal smoothing and DiffPure.
Experimental setup: On the complete benchmarks (Sintel, Bonn, KITTI), three temporal filters are applied to the protected videos: a two-frame average and two motion-aligned three-frame medians (Table 2 in the manuscript). Following your suggestion, we conduct two more experiments on one 50-frame Sintel clip (alley_1): a centered five-frame average, and frame-by-frame DiffPure [d], which adds noise to each frame and removes it with a diffusion model (noise timestep 150). DiffPure runs at 256 × 256, so we also report a resize-only control that applies the same down- and up-sampling without diffusion.
Results: Under the three temporal filters, Safe4D keeps 91.3–108.3% of its depth-error increase on the three datasets. On alley_1 (Table C2; Tables 33 and 35 in the revised manuscript), the five-frame average keeps 70.7%.
Table C2: Attacker-side processing on all 50 frames of alley_1 (AbsRel, the absolute relative depth error of StreamVGGT). Bracketed values indicate changes relative to the clean video under the same processing, computed from unrounded scores.
| Processing | Clean AbsRel (↑) | Protected AbsRel (↑) |
|---|---|---|
| None (direct) | 0.173 | 0.844 (+0.671) |
| Centered five-frame average | 0.212 | 0.686 (+0.474) |
| Resize only | 0.240 | 0.286 (+0.0462) |
| DiffPure | 0.274 | 0.275 (+0.000186) |
(2) The +1.94 dB result. This number comes from a supplementary experiment in the appendix (Table 25 in the submission, Table 27 in the revised manuscript). It reconstructs one scene with 4D Gaussian splatting, with one protected camera out of 20, and it uses the earlier Same-grid variant (which copies the previous perturbation to the same pixels) instead of the final Safe4D. In that run, the answer is yes: when the attacker skips the filtering, the rendered held-out views are not degraded. Their PSNR (peak signal-to-noise ratio; higher means better rendering) is 1.94 dB higher than with the clean input, and after two-frame averaging it is 2.23 dB lower. For depth, which is what Safe4D optimizes, skipping the filter does not help the attacker: without any filtering, the depth error is already 1.96–3.66× that of the clean video (Table 1).
References
[d] Diffusion Models for Adversarial Purification.
Response to Reviewer 9Df6 (Part 3/5)
W2 & Q3: Could an amortized generator reduce the per-frame cost by two to three orders of magnitude?
Response 3: Thank you for the constructive suggestion. Following your suggestion, we conduct a preliminary experiment that trains a small network to output the perturbation in one forward pass, instead of optimizing every frame iteratively.
Experimental setup: The generator is a small convolutional network (three 3 × 3 convolution layers, about 11k parameters). It takes the current frame as input and is trained to reproduce the perturbation that iterative Safe4D produces for that frame. It is trained for 300 updates on the first 16 frames of one Sintel clip (alley_2) and tested on all 50 frames of another clip (alley_1) that it has never seen.
Results: The results are reported in Table C3 (Table 34 in the revised manuscript). The generator produces a perturbation in 0.280 ms per frame, but this perturbation keeps only 0.46% of the depth-error increase of iterative Safe4D The network here is very small and is trained for only 300 updates on a single clip. Training and evaluating a generator on the full datasets takes longer time, and we will report those results in the final version.
Table C3: One-pass generator vs. iterative Safe4D on the held-out clip alley_1. The two timings are measured differently and are not an end-to-end comparison: 0.280 ms covers network inference, clipping and rounding on one H200 GPU, and 291 s is the median over the 70 benchmark videos.
| Method | Time per frame (↓) | Protected AbsRel (↑) | ΔAbsRel (↑) | Protection kept (↑) |
|---|---|---|---|---|
| Iterative Safe4D | 291 s | 0.844 | +0.671 | 100% |
| One-pass generator | 0.280 ms | 0.176 | +0.0031 | 0.46% |
Response to Reviewer 9Df6 (Part 1/5)
Dear Reviewer
9Df6,Thanks for your constructive comments! We respond to each of your points below. Please feel free to let us know if further details or explanations are helpful.
In all tables of this response, ↑ (↓) after a metric means that a higher (lower) value is better from the protector's side. For example, a higher depth error means stronger protection.
Response 1: Thank you for the detailed comments. We answer the three parts in turn.
(1) Full distribution. The submission reports every sequence for CUT3R [c], a streaming model with a different architecture and training data, in Appendix G (Tables 20–23 in the revised manuscript). Following your suggestion, we add two more unseen models, VGGT [a] (offline) and STream3R [b] (streaming), report every sequence for them as well (Table 37 in the revised manuscript), and summarize the three models in Table C1 (Table 5 in the revised manuscript). Safe4D is optimized against StreamVGGT only, and the three models receive the same protected videos without retuning. On the released videos, the depth error increases on 61/62 sequences for VGGT and 62/62 for STream3R, and on 34/62 for CUT3R.
(2) Why transfer sometimes fails. The failures concentrate on CUT3R. VGGT and STream3R are related to StreamVGGT, and transfer to them is consistent. CUT3R has a different architecture and training data: the perturbation still changes its prediction, but without a consistent direction. On the 23 Sintel sequences, the median change in its depth error is +1.52 × 10⁻³, and the middle half of the sequences lies between −23.8 × 10⁻³ and +36.1 × 10⁻³.
Table C1: Depth transfer on all 62 depth sequences. ΔAbsRel is the absolute relative depth error on the protected video minus that on the clean video under the same processing (a positive value means worse reconstruction). Each cell gives the mean ΔAbsRel over the sequences in units of 10⁻³ and, in brackets, the number of sequences with a higher error. "Forward two-frame average" means that the attacker averages every frame with the next one.
References
[a] VGGT: Visual Geometry Grounded Transformer.
[b] STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer.
[c] Continuous 3D Perception Model with Persistent State.