acceptodds
Under review as a conference paper at ICLR 2027

Meaningful Anonymization for Privacy-Preserving Video Generation

Abstract

Videos contain rich visual and temporal information for understanding human activities, but can also expose sensitive information about people and their surroundings. Existing privacy-preserving video methods primarily reduce such leakage by suppressing or discarding privacy-sensitive visual information, which can inadvertently remove content useful for understanding the video. We introduce **Mask2Safe**, a privacy-preserving video generation framework that learns *where* privacy-preserving modifications are needed and uses this localization to guide *how* sensitive content is transformed. Rather than relying on the costly vision-language reasoning for video frames, Mask2Safe learns semantic privacy localization from paired unsafe–safe images through a lightweight *Privacy Masker*. Its predictions are refined and propagated over the whole clip to produce temporally consistent masks that support either mask-based or generative anonymization. We further introduce **SafeQA**, a video question-answering benchmark that evaluates meaningful anonymization along three complementary dimensions: *Privacy*, whether the original sensitive information remains recoverable; *Visibility*, whether the affected visual content remains observable; and *Local Utility*, whether nearby non-sensitive information remains recoverable. Experiments on VP-UCF101 and VP-HMDB51 show that Mask2Safe reduces the predictability of sensitive attributes while retaining downstream utility on action recognition task, and that classifiers trained on our anonymized videos are able to transfer effectively to the original video domain. SafeQA further shows that generative anonymization substantially improves visibility over mask-based anonymization while maintaining strong local utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
privacy preservation, action recognition, video understanding

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_DNCqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Mask2Safe to remove identity and sensitive attributes while keeping other relevant information in the video. Mask2Safe trains a lightweight Privacy Masker (frozen DINOv2 + small decoder) on unsafe/safe image pairs from Unsafe2Safe. It learns which regions must change by using a U-Net to predict replacement content. On video, the per-frame masks seed SAM2, which refines and tracks the region within each shot. The region is then either blacked out or regenerated with a video inpainting model. The paper also proposes SafeQA, a benchmark to tell apart if privacy is achieved by hiding content (visibility) or is achieved by changing only the sensitive parts (local utility).

Critical strengths

The paper explores replacement instead of suppression, with real-video transfer. Mask2Safe edits only localized regions so classifiers trained on its videos transfer to original videos, while Unsafe2Safe drops almost 20%. SafeQA probes the edited regions themselves, which clip-level metrics cannot do. This is a useful direction.

Critical weaknesses

The mask-learning method is not shown to be needed. The trainable U-Net (later discarded) and the L_pref and L_preserve losses are never ablated. Simpler mask sources are never compared: thresholded source and target image differences, a frozen pretrained editor, or detector + SAM2 person masks. This matters because refined masks often cover whole people (appendix figure 12). If a person segmenter performs as well, the claim that image-learned privacy knowledge transfers to video is unsupported.

Furthermore, cMAP improves only from 72.86 to 71.07 on VP-UCF101 (4× downsampling: 71.61). On SafeQA Privacy, Mask2Safe is worse than Unsafe2Safe and TeD-SPAD. SafeQA relies on a single Gemini judge with no human validation and has about 20% chance accuracy on Q1 and Q3. SafeQA is the paper’s main evidence for the benefit of replacement and this undermines the central claim.

Questions for authors

The refined masks in Figs. 11 and 12 often cover whole people. How do the learned masks compare with detector + SAM2 person masks, or with thresholded source-target latent differences used as pseudo-labels, under the same black-out and EditCtrl inpainting?

The U-Net G is discarded after training, and L_pref and L_preserve mainly train G rather than the mask. Are they needed? How do the masks change with a frozen pretrained editor (for instance, InstructPix2Pix) as G, or without L_pref, L_preserve or the stop-gradients?

Appendix A.2 prompts EditCtrl with “A video of {class name}”. Is the ground-truth label also used for test videos? If so, what is the action accuracy in Table 1 with a class-agnostic prompt?

SafeQA questions are generated and answered by Gemini models with no human check. Could you report human agreement on a random subset and results with a second (ideally open-source) judge?

7 comments
AuthorsOpenReview ↗

We thank the reviewers for the great suggestions and constructive feedbacks. Response to Weakness 1

We thank the reviewer for recognizing the value of localized replacement, improved anonymized-to-original transfer, and SafeQA's fine-grained privacy-utility evaluation. We appreciate the constructive questions and provide additional experimental evidence and methodological clarifications below along with the revised paper that will be updated shortly.

W1. Necessity of Learned Privacy Localization

Following the reviewer's suggestions, we provide ablations of the Privacy Masker and comparisons with simpler pseudo-label supervision and pretrained person segmentation (Full details in Appendix E, Tables 5-6, Figures 13-15).

Training objective ablations (Table 5).
We ablate the regional reconstruction losses, area regularization, and stop-gradient operations when training the Privacy Masker on Unsafe2Safe editing pairs. We also compare against models trained with differences between the unsafe-safe pairs on pixel level and DINO feature level as pseudo-targets. All variants are evaluated off-the-shelf against ground-truth privacy annotations on BIV-Priv-HIT (87 annotated videos), used here only as a reference given its mismatched privacy taxonomy and our coarse resolution masks.

Privacy Masker ablations and pseudo-label baselines on BIV-Priv-HIT (Appendix Table 5). Higher sensitive-part AUROC and Best F1 are better; [email protected] measures predicted mask coverage at threshold τ=0.5\tau=0.5.

VariantAUROC ↑\uparrowBest F1 ↑\uparrow[email protected]
Mask2Safe (full)0.8960.5079.5%
Without Lpref\mathcal{L}_{\mathrm{pref}}, no area penalty0.8690.45730.2%
Without Lpref\mathcal{L}_{\mathrm{pref}}, with area penalty0.6590.2450.0%
Without Lpreserve\mathcal{L}_{\mathrm{preserve}}, no area penalty0.6410.2260.0%
Without U-Net stop-gradient0.8900.5029.3%
Without all stop-gradients0.5150.17312.0%
DINO-difference pseudo-target0.4880.1739.1%
Pixel-difference pseudo-target0.6200.2110.0%

The auxiliary U-Net GG provides a training-time reconstruction pathway that encourages the Privacy Masker to identify regions requiring replacement without spatial supervision. Although LprefL_{\mathrm{pref}} and LpreserveL_{\mathrm{preserve}} supervise this pathway, their removal directly affects the learned masks. Without LprefL_{\mathrm{pref}}, masks become overly extensive in the absence of area regularization or collapse when it is retained. Removing Lpreserve\mathcal{L}_{\mathrm{preserve}} also substantially degrades localization. Removing only the U-Net stop-gradient introduces artifacts on the token borders, whereas removing all stop-gradients collapses into almost black masks. These results demonstrate the importance of the regional reconstruction objectives and gradient constraints, without which the Privacy Masker will result in a collapse mode or an imperfect mask.

Both pseudo-label strategies perform substantially worse than the full model, indicating that simple label construction is insufficient to guide the learning of privacy localization. We found that such differences are sensitive to noise and changes unrelated to privacy. Additional implementation details and qualitative comparisons that visualize the artifacts are provided in Appendix E.

Comparison with object segmentation (Table 6). On COCO and Open Images, we evaluate class-wise pixel AUROC (distinguishing privacy-related objects from surroundings) and the selectivity gap (mean AUROC difference between privacy-related and non-sensitive control classes) against a person detector based on Mask R-CNN backbone.

Privacy discrimination and selectivity compared with pretrained object segmentation (Appendix Table 6). Higher privacy AUROC and selectivity gap are better.

MethodCOCO Privacy AUROC ↑COCO Selectivity Gap ↑OpenImages Privacy AUROC ↑OpenImages Selectivity Gap ↑
Mask R-CNN (person only)0.5480.0850.5470.094
Mask2Safe0.7430.3360.6980.294

Compared with person-only Mask R-CNN, which primarily responds to people, our Privacy Masker assigns substantially higher mask intensities to diverse privacy-sensitive objects than to non-sensitive categories. As shown in Appendix E.3 and Figure 15, mask intensity inside privacy-related objects is 2.24 times that of their surroundings on COCO (1.97 times on Open Images), compared with only 0.29 and 0.28 for non-sensitive objects, respectively. These results demonstrate that our learned masks capture privacy relevance beyond person detection.

Although SAM2 can propagate person masks across video frames, it does not resolve the limitations restricting to a predefined class and no semantic adapting. Our results, in contrast, support learning privacy-selective spatial representations from image-level supervision.

AuthorsOpenReview ↗

Response to Reviewer DNCq's Weakness 2 and Questions 1-2

W2. Privacy–Utility Trade-off and SafeQA Reliability (Appendices G–H)

We agree that the reduction in privacy cMAP is modest. However, cMAP measures the recoverability of predefined privacy attributes and does not fully characterize other forms of privacy leakage, such as facial identity linkability.

Thus, we provide an additional evaluation of cross-video identity leakage using ArcFace embeddings (Appendix G). Specifically, we measure facial similarity across different videos containing the same person. Ideally, anonymization should make it difficult to recognize that the same person appears in multiple videos. On VP-UCF101 and VP-HMDB51, cross-video facial similarity (lower is better) decreases from 0.769 to 0.223 and from 0.722 to 0.224, respectively, after generative anonymization. These results indicate substantially reduced identity linkability across videos.

More importantly, the benefit of our generative replacement is supported by SafeQA. Under the same privacy localization pipeline, replacing sensitive regions rather than directly masking them improves original-domain action accuracy by 9.49 and 5.03 percentage points on VP-UCF101 and VP-HMDB51, respectively. Compared with Unsafe2Safe, our method also improves original-domain action accuracy by 11.18 and 1.69 points (Table 1). These results directly support the utility of generative replacement for downstream video understanding.

To assess evaluator dependence, we extend SafeQA evaluation to three open-source VLMs: Qwen3.5-4B, Qwen3.5-9B, and Qwen3-VL-8B (Appendix H). We examine whether the conclusions regarding privacy recovery and local utility remain consistent. We report representative Gemini and Qwen3.5-4B results in the rebuttal for brevity, with additional evaluations provided in the appendix.

SafeQA results evaluated using Gemini and Qwen3.5-4B. Bold and italic indicate the best and second-best results, respectively, among anonymization methods.

MethodGemini Q1 ↓Gemini Q2 ↑Gemini Q3 ↑Qwen Q1 ↓Qwen Q2 ↑Qwen Q3 ↑
VP-UCF101
Raw97.398.194.390.096.392.8
Downsample 2×91.297.391.484.293.990.6
Downsample 4×78.495.687.474.980.486.2
TeD-SPAD55.068.866.460.621.165.9
Unsafe2Safe50.671.765.960.172.570.2
Mask2Safe (Mask-based)62.350.674.964.825.873.0
Mask2Safe (Generative)63.877.173.668.174.677.5
VP-HMDB51
Raw97.998.594.688.191.790.6
Downsample 2×84.286.785.584.687.388.4
Downsample 4×74.286.882.176.776.583.8
TeD-SPAD58.465.568.960.824.270.0
Unsafe2Safe49.172.064.660.669.069.2
Mask2Safe (Mask-based)62.047.273.364.423.374.7
Mask2Safe (Generative)71.080.377.371.177.079.9

We are extending SafeQA with human evaluation on the question quality and will report these results once completed.

These findings reinforce the utility benefits of generative replacement. That said, stronger privacy suppression remains an area for improvement.

Q1. Comparison Under Identical Anonymization Pipelines

As detailed in W1, our additional experiments demonstrate improved privacy localization over difference-based pseudo-labels and pretrained person segmentation.

Q2. Auxiliary U-Net, Regional Losses, and Stop-Gradient

W1 also provides ablations showing that the regional reconstruction losses are important for learning meaningful privacy masks, despite the auxiliary U-Net being discarded after training. The individual U-Net stop-gradient has limited impact, whereas removing all stop-gradients substantially degrades localization.

AuthorsOpenReview ↗

Response to Reviewer DNCq's Questions 3-4

Q3. Action-Conditioned Versus Class-Agnostic Editing (Appendix C.3, F.2-F.6, Table 11)

In the main evaluation, we have two different test settings for utility: (1) the accuracy on anonymized videos and (2) the accuracy on original test videos. We clarify that action-conditioned prompts are used to generate anonymized videos for training and the first setting, but not for the second one (original-domain evaluation). In the second setting, classifiers are evaluated on unmodified original videos, without editing or ground-truth action prompts at inference. Therefore, the original-domain results cannot be attributed to action-conditioned test-video generation.

To directly assess potential prompt-induced bias in same-domain evaluation, we additionally evaluate class-agnostic editing on the test videos. Instead of supplying the ground-truth action class, we use per-video edit instruction generated by Gemini. These descriptions retain observable movements, object interactions, and scene structure while modifying identifying appearance without mentioning the class name.

MethodTest VideoEditing PromptVP-UCF101 ↑VP-HMDB51 ↑
Mask2Safe (Ours)Original--76.5139.08
AnonymizedAction-conditioned74.5339.93
AnonymizedClass-agnostic (Gemini)72.1438.82
TeD-SPADAnonymized--62.5830.65

Action-recognition accuracy (Top-1, %, ↑) under different test-time editing settings. Full results are reported in Appendix Table 11.

Removing ground-truth action prompts reduces same-domain accuracy by only 2.39 percentage points on VP-UCF101 and 1.11 points on VP-HMDB51. Even without action-label prompting, Mask2Safe achieves 72.14% and 38.82%, exceeding TeD-SPAD by 9.56 and 8.17 percentage points, respectively.

Test-set prompt templates are given in Appendix C.3; evaluation protocols and results are in Appendix F. We will update Table 1 and the main paper (highlighted in blue) to integrate this test score. Our evaluation conclusions remain unchanged.

Q4. Independent Judge and Human Agreement

As discussed in Response 2, evaluation with open-source VLM judges yields consistent conclusions regarding SafeQA privacy and utility. We will report human evaluation on the generated benchmark questions once completed.

Reviewer_55pgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Mask2Safe, a privacy-preserving video anonymization framework that aims to remove sensitive visual information while preserving non-sensitive content useful for downstream understanding.

The method learns privacy-sensitive region localization from paired unsafe–safe images using a lightweight Privacy Masker. It then applies SAM2 for temporal mask refinement and propagation, followed by either mask-based anonymization or generative replacement using a pretrained video editor.

The authors additionally introduce SafeQA, a video question-answering benchmark that evaluates anonymization along three dimensions: privacy protection, visibility of modified content, and preservation of nearby non-sensitive visual information.

Experiments on VP-UCF101 and VP-HMDB51 demonstrate promising results in maintaining downstream action recognition and reducing anonymized-to-original domain shift

Critical strengths
  • Fine-grained privacy–utility evaluation. SafeQA provides an interesting evaluation perspective by separately measuring privacy, visibility, and local utility. This extends conventional evaluations based primarily on privacy-attribute prediction and action recognition.

  • Practical framework design. Transferring privacy localization learned from images to videos is a reasonable approach to reducing repeated frame-level VLM inference. The integration with SAM2 and generative editing provides a practical video anonymization pipeline.

  • Promising cross-domain transfer results. The anonymized-to-original transfer experiments suggest that localized transformations can better preserve task-relevant information than methods introducing substantial appearance changes across entire frames.

Critical weaknesses
  • The paper measures privacy leakage using predefined privacy-attribute classifiers and VLM-based question answering. While these evaluations are useful, reduced prediction accuracy does not necessarily imply that sensitive information has been successfully removed. For instance, an anonymized video might prevent a particular classifier from recognizing an attribute while still retaining sufficient information for another model or human observer to recover it. This concern becomes more important with generative anonymization, because modified appearance does not necessarily prevent re-identification through temporal context, clothing, background, or other correlated visual features. Stronger experiments could include identity re-identification, cross-frame information aggregation, and privacy classifiers trained specifically to recover sensitive information from anonymized videos.

  • Temporal consistency is a central motivation but is not directly evaluated. A major motivation for extending image-level anonymization to video is the need to maintain consistent privacy-preserving transformations across time. The proposed method uses SAM2 for temporal mask propagation and a video editor for temporally coherent replacement. However, temporal consistency is evaluated primarily through qualitative examples rather than dedicated quantitative metrics. SafeQA also focuses mainly on content-level preservation rather than temporal associations. The paper would benefit from quantitative evaluations of mask consistency, temporal flickering, identity consistency of generated replacements, and privacy leakage accumulated across multiple frames. This is particularly important because temporal consistency is presented as a central advantage over independently anonymizing individual frames.

Questions for authors
  • What is the computational efficiency?. Although the Privacy Masker is described as lightweight, the complete pipeline includes shot detection, SAM2 propagation, and generative video editing. End-to-end inference cost and runtime comparisons would strengthen the practical claims.

  • The anonymized-to-original transfer metric requires careful interpretation. A small absolute accuracy gap does not necessarily indicate good transfer if both accuracies are low. The transfer metric should be interpreted alongside absolute action-recognition performance.

  • Failure-case analysis is limited. Additional examples involving missed sensitive regions, incorrect localization, rapid camera movement, occlusion, and generative artifacts would help characterize the method's robustness.

5 comments
AuthorsOpenReview ↗

We thank the reviewer for the careful review and helpful feedback. Reponse 1/3

We thank the reviewer for recognizing the strengths of SafeQA’s fine-grained privacy-utility evaluation, our image-to-video anonymization framework, and the promising original-domain generalization results. We provide methodological clarifications and additional quantitative evidence addressing the reviewer’s concerns below along with the revised paper that will be updated shortly.

W1. Privacy Leakage Evaluation and Recovery Attacks (Appendices F-H)

We agree that reduced privacy classifier accuracy alone does not establish complete removal of sensitive information. In fact, our evaluation already examines attribute recovery under domain adaptation and privacy recovery by robust VLMs benchmark. In the revised Appendix G, we also evaluate cross-frame identity similarity.

First, privacy classifiers are trained separately for each anonymization method; we did not evaluated only with a fixed classifier trained on original videos. Following the STPrivacy evaluator setting, we train a ViT-S video model on each method's anonymized training data to predict the five binary privacy attributes. Each model takes 16-frame clips, and test predictions are averaged over five temporal views and three spatial crops. Because the classifier is trained on the anonymized videos, it can exploit residual privacy cues and method-specific artifacts.

Second, we provide an additional evaluation of cross-video identity leakage using ArcFace embeddings (Appendix G). Specifically, we measure facial similarity across different videos containing the same person. Ideally, anonymization should make it difficult to recognize that the same person appears in multiple videos. On VP-UCF101 and VP-HMDB51, cross-video facial similarity (lower is better) decreases from 0.769 to 0.223 and from 0.722 to 0.224, respectively, after generative anonymization. These results indicate substantially reduced identity linkability across videos.

Third, in this work we focused on the SafeQA benchmark whose questions having target-specific, visually grounded privacy attributes and nearby non-sensitive information, rather than relying on global appearance or low-level visual statistics. This makes the benchmark applicable to both our realistic generation and pixel-distorting approaches such as TeD-SPAD and SPAct. To assess evaluator dependence, we extend SafeQA evaluation to three open-source VLMs: Qwen3.5-4B, Qwen3.5-9B, and Qwen3-VL-8B (Appendix H). We examine whether the conclusions regarding privacy recovery and local utility remain consistent. We report representative Gemini and Qwen3.5-4B results in the rebuttal for brevity, with additional evaluations provided in the appendix.

As indicated in the table below, both evaluators consistently give higher Q3 Local Utility for Mask2Safe (Generative), but also higher Q1 privacy recovery than Unsafe2Safe and TeD-SPAD.

SafeQA results evaluated using Gemini and Qwen3.5-4B. Bold and italic indicate the best and second-best results, respectively, among anonymization methods.

MethodGemini Q1 ↓Gemini Q2 ↑Gemini Q3 ↑Qwen Q1 ↓Qwen Q2 ↑Qwen Q3 ↑
VP-UCF101
Raw97.398.194.390.096.392.8
Downsample 2×91.297.391.484.293.990.6
Downsample 4×78.495.687.474.980.486.2
TeD-SPAD55.068.866.460.621.165.9
Unsafe2Safe50.671.765.960.172.570.2
Mask2Safe (Mask-based)62.350.674.964.825.873.0
Mask2Safe (Generative)63.877.173.668.174.677.5
VP-HMDB51
Raw97.998.594.688.191.790.6
Downsample 2×84.286.785.584.687.388.4
Downsample 4×74.286.882.176.776.583.8
TeD-SPAD58.465.568.960.824.270.0
Unsafe2Safe49.172.064.660.669.069.2
Mask2Safe (Mask-based)62.047.273.364.423.374.7
Mask2Safe (Generative)71.080.377.371.177.079.9
AuthorsOpenReview ↗

Response to Reviewer 55PG WQBh 2/3

W2. Quantitative Temporal Consistency (Appendices E-G)

Following the reviewer's suggestions, we evaluate temporal consistency of different stages of our pipeline through mask IoU, flicker rate, and facial appearance stability.

Temporal mask consistency (Appendix E.4, Table 7).

We apply optical flow warping to the masks and measure IoU between frames in all VP-UCF101 and VP-HMDB51 clips, and found that the mask refined by SAM2 is more temporally coherent than the coarse Privacy Masker predictions. Under fast motion, the IoU decreases by only 0.094 for SAM2 versus 0.213 for coarse masks.

Temporal mask consistency before and after SAM2 refinement. Higher flow-warped IoU is better.

DatasetCoarse MaskSAM2-Refined Mask
VP-UCF1010.6570.885
VP-HMDB510.6640.865

Edited video flickering

We evaluate temporal smoothness on the full test splits using the VBench[R1] temporal flickering score, computed from the pixel similarity between consecutive video frames (higher is better). The results suggest that thanks to the generative step, our anonymized output maintains temporal coherence comparable to the original videos, regardless of whether coarse (without SAM2) or refined (with SAM2) privacy masks are used.

MethodSAM2 RefinementVP-UCF101 ↑VP-HMDB51 ↑
Original video—97.8997.85
Mask2Safe (Generative)✗97.8997.84
Mask2Safe (Generative)✓97.9297.79

Generated appearance consistency (Appendix G, Table 12).

To assess whether video-level anonymization maintains consistent facial appearances across frames, rather than introducing the identity fluctuations associated with independent framewise editing, we compute within-video ArcFace embedding similarity [R2] across matched face detections from six sampled frames per clip:

Within-video facial appearance consistency measured by ArcFace cosine similarity. Higher is better.

DatasetFramewise Unsafe2SafeVideo-Level Mask2Safe
VP-UCF1010.3130.599
VP-HMDB510.3050.578

In addition to maintaining a better temporal flickering score, video-level generation produces substantially more consistent facial appearances across frames than framewise editing. Meanwhile, the low similarity to original faces (0.214 on VP-UCF101 and 0.195 on VP-HMDB51) indicates that this temporal consistency does not come at the expense of preserving the original facial identity.

[R1] Huang, Ziqi, et al. "Vbench: Comprehensive benchmark suite for video generative models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

[R2] Deng, Jiankang, et al. "ArcFace: Additive Angular Margin Loss for Deep Face Recognition." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

AuthorsOpenReview ↗

Response to Reviewer 55PG 3/3

Q1. Computational Efficiency

We thank the reviewer for raising this question. We measured end-to-end inference latency on a single A100-80GB GPU at 240×320240\times320 resolution, processing one clip at a time.

ClipPrivacy MaskerSAM2EditCtrlTotal
UCF101 (186 frames, ~7.5 s)0.04 s12.3 s21.9 s~34 s
HMDB51 (94 frames, ~3.1 s)0.02 s6.6 s11.9 s~18 s

As shown in the table, the Privacy Masker itself is extremely fast (0.02-0.04 s per clip), and total latency is dominated by SAM2 propagation and EditCtrl generation, partially because both adapted to our implementation without optimized batch processing. These modules was selected as the mask refiner and generative editor in our framework to demonstrate its effectiveness and can be replaced with more efficient alternatives depending on deployment and anonymization requirements.

The primary computational advantage of Mask2Safe lies in replacing the repeated VLM privacy inspection of Unsafe2Safe with a lightweight, feed-forward Privacy Masker, which directly predicts sensitive regions. While SAM2 and EditCtrl introduce additional costs, anonymization can be performed offline to prepare data for training downstream models.

Q2. Interpretation of Anonymized-to-Original Transfer (Main Paper Table 1)

We agree that a small transfer gap alone does not establish strong task utility and accordingly Table 1 has reported both absolute action-recognition accuracy and the anonymized-to-original transfer gap.

Mask2Safe (Generative) achieves 76.51% original-domain accuracy on VP-UCF101, compared with 65.33% for Unsafe2Safe (+11.18 points), and 39.08% versus 37.39% on VP-HMDB51 (+1.69 points).

On VP-UCF101, Unsafe2Safe drops from 85.14% on anonymized videos to 65.33% on original videos, whereas Mask2Safe achieves 74.53% and 76.51%, respectively. Thus, Mask2Safe achieves higher absolute original-domain accuracy alongside a smaller transfer gap, although Unsafe2Safe remains stronger on anonymized-domain recognition.

Q3. Failure Cases and Robustness (Appendix I, Figure 17)

Thanks to the reviewer's recommandation, we extend the failure-case analysis in Appendix I to examine limitations in privacy localization, temporal mask propagation, and generative editing.

Localization sensitivity. Sometimes the masker might over-mask (covering more than the attributes defined by the privacy criteria) (Figure 17).

Occlusion and rapid motion. SAM2 propagation can introduce mask drift onto background regions, particularly during occlusion and rapid motion, despite improved overall temporal stability.

Generative artifacts. Editing within a mask may introduce texture changes, especially for large masked regions, and might preserve unsafe content outside the edited areas.

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

The paper introduces Mask2Safe, a privacy-preserving video generation framework that learns privacy-sensitive regions from paired unsafe–safe images. It uses a lightweight Privacy Masker, SAM2-based temporal refinement, and generative editing to replace sensitive content. The paper also introduces SafeQA, a benchmark evaluating privacy, visibility, and local utility through video question answering.

Critical strengths
  1. The paper addresses the privacy–utility trade-off by replacing sensitive content rather than simply suppressing visual information.
  2. Mask2Safe transfers privacy localization from images to videos without requiring annotated privacy masks or paired anonymized videos.
  3. SafeQA introduces three complementary dimensions: Privacy, Visibility, and Local Utility. This provides a useful perspective beyond conventional action-recognition accuracy.
Critical weaknesses
  1. Generative anonymization uses the ground-truth action class in its editing prompt in Appendix A.2. This may bias action-recognition results. Evaluation with class-agnostic prompts is needed.

  2. SafeQA assigns "Yes" as the ground-truth answer to every visibility question in Appendix C.2. The reported accuracy therefore reflects the VLM's frequency of answering "Yes". Independent validation may improve the paper's fairness.

  3. Mask2Safe provides modest reductions in privacy cMAP. Its generative variant also shows higher SafeQA privacy recovery than Unsafe2Safe on both datasets (Tables 1–2).

  4. Generative Mask2Safe achieves lower action accuracy than the raw baseline and Unsafe2Safe on both datasets in Table 1. A small domain-transfer gap alone does not establish better task utility.

  5. The contributions of privacy localization, SAM2 refinement, and generative editing are not independently evaluated. Quantitative mask quality and temporal consistency analyses are also missing.

  6. The results are not compared with some important papers. Ex. "Balancing Privacy and Action Performance: A Penalty-Driven Approach to Image Anonymization".

  7. The author should follow a ranking method to rank the action-privacy table for readers' easy understanding, as the results comprise three different metrics where two (cMAP, F1) should be minimized and one should be maximized (action performance).

Questions for authors
  1. How does Mask2Safe perform when generative editing uses class-agnostic prompts instead of ground-truth action labels?
  2. Can the authors provide component-wise ablations and quantitative temporal consistency results to establish the effectiveness of the proposed pipeline?
6 comments
AuthorsOpenReview ↗

We appreciate the reviewer's detailed feedbacks and suggestions. Response 1/4

We thank the reviewer for recognizing the contributions of localized generative anonymization, annotation-free image-to-video privacy localization, and SafeQA's complementary evaluation of privacy and visual utility. We provide additional experiments and methodological clarifications addressing the concerns below, along with the revised paper that will be updated shortly.

W1. Action-Conditioned Editing and Recognition Bias (Appendices C.3 and F)

In the main evaluation, we have two different test settings for utility: (1) the accuracy on anonymized videos and (2) the accuracy on original test videos. We clarify that action-conditioned prompts are used to generate anonymized videos for training and the first setting, but not for the second one (original-domain evaluation). In the second setting, classifiers are evaluated on unmodified original videos, without editing or ground-truth action prompts at inference. Therefore, the original-domain results cannot be attributed to action-conditioned test-video generation.

To directly assess potential prompt-induced bias in same-domain evaluation, we additionally evaluate class-agnostic editing on the test videos. Instead of supplying the ground-truth action class, we use per-video edit instructions generated by Gemini. These descriptions retain observable movements, object interactions, and scene structure while modifying identifying appearance without mentioning the class name.

Action-recognition accuracy (Top-1, %, ↑) under different test-time editing settings. Full results are reported in Appendix Table 11.

MethodTest VideoEditing PromptVP-UCF101 ↑VP-HMDB51 ↑
Mask2Safe (Ours)Original--76.5139.08
AnonymizedAction-conditioned74.5339.93
AnonymizedClass-agnostic (Gemini)72.1438.82
TeD-SPADAnonymized--62.5830.65

Removing ground-truth action prompts reduces same-domain accuracy by only 2.39 percentage points on VP-UCF101 and 1.11 points on VP-HMDB51. Even without action-label prompting, Mask2Safe achieves 72.14% and 38.82%, exceeding TeD-SPAD by 9.56 and 8.17 percentage points, respectively.

Test-set prompt templates are given in Appendix C.3; evaluation protocols and results are in Appendix F. We will update Table 1 and the main paper to integrate this test score. Our evaluation conclusions remain unchanged.

AuthorsOpenReview ↗

Response to Reviewer WQBh 2/4

W2. SafeQA Visibility Evaluation and Affirmative-Response Bias (Appendix H)

We clarify that SafeQA Q2 is designed to measure whether originally visible, utility-relevant content remains observable after anonymization, rather than general binary visibility classification. Accordingly, every Q2 question is expected to have positive reference answers. Yet, it may be susceptible to affirmative-response bias, particularly when VLMs encounter out-of-distribution anonymized inputs.

To assess whether our comparative findings are sensitive to the choice of evaluator, we extend SafeQA evaluation to three open-source VLMs: Qwen3.5-4B, Qwen3.5-9B, and Qwen3-VL-8B (Appendix H). We examine whether the conclusions regarding privacy recovery and local utility remain consistent. We report representative Gemini and Qwen3.5-4B results in the rebuttal for brevity, with additional evaluations provided in the appendix.

As shown in the table below, both evaluators consistently show that generative replacement improves Q2 Visibility over direct masking. Generative Mask2Safe also outperforms Unsafe2Safe in Q3 Local Utility on both datasets, which aligns with our conclusion in the main paper.

SafeQA results evaluated using Gemini and Qwen3.5-4B. Bold and italic indicate the best and second-best results, respectively, among anonymization methods.

MethodGemini Q1 ↓Gemini Q2 ↑Gemini Q3 ↑Qwen Q1 ↓Qwen Q2 ↑Qwen Q3 ↑
VP-UCF101
Raw97.398.194.390.096.392.8
Downsample 2×91.297.391.484.293.990.6
Downsample 4×78.495.687.474.980.486.2
TeD-SPAD55.068.866.460.621.165.9
Unsafe2Safe50.671.765.960.172.570.2
Mask2Safe (Mask-based)62.350.674.964.825.873.0
Mask2Safe (Generative)63.877.173.668.174.677.5
VP-HMDB51
Raw97.998.594.688.191.790.6
Downsample 2×84.286.785.584.687.388.4
Downsample 4×74.286.882.176.776.583.8
TeD-SPAD58.465.568.960.824.270.0
Unsafe2Safe49.172.064.660.669.069.2
Mask2Safe (Mask-based)62.047.273.364.423.374.7
Mask2Safe (Generative)71.080.377.371.177.079.9

We are extending SafeQA with human evaluation on the question-answer quality and will report these results once completed.

W3. Privacy Recovery Versus Meaningful Visual Utility

We acknowledge that Mask2Safe does not achieve the strongest privacy suppression across all metrics. Our objective is to balance privacy protection with preservation of action-relevant visual information, rather than maximize suppression alone.

In the table we showed in W2, with both independent VLM judges, Mask2Safe (Generative) consistently improves Q2 Visibility and Q3 Local Utility over Unsafe2Safe, albeit with higher Q1 privacy recovery. In addtion to SafeQA, Mask2Safe (Generative) improves original-domain action recognition over the Mask2safe (Mask-based) variant by 9.49/5.03 points (67.02% → 76.51% on UCF101; 34.05% → 39.08% on HMDB51). These findings reinforce the utility benefits of generative replacement. That said, stronger privacy suppression remains an area for improvement.

W4. Action Recognition and Original-Domain Generalization

We agree that a small transfer gap alone does not establish strong task utility and, accordingly, Table 1 has reported both absolute action-recognition accuracy and the anonymized-to-original transfer gap.

Mask2Safe (Generative) achieves 76.51% original-domain accuracy on VP-UCF101, compared with 65.33% for Unsafe2Safe (+11.18 points), and 39.08% versus 37.39% on VP-HMDB51 (+1.69 points).

On VP-UCF101, Unsafe2Safe drops from 85.14% on anonymized videos to 65.33% on original videos, whereas Mask2Safe achieves 74.53% and 76.51%, respectively. Thus, Mask2Safe achieves higher absolute original-domain accuracy alongside a smaller transfer gap, although Unsafe2Safe remains stronger on anonymized-domain recognition.

AuthorsOpenReview ↗

Response to Reviewer WQBh 3/4

W5. Contributions of Privacy Localization, SAM2, and Generative Editing (Appendices E-G)

Following the reviewer's suggestions, we evaluate the temporal consistency of different stages of our pipeline through mask IoU, flicker rate, and facial appearance stability.

Temporal mask consistency (Appendix E.4, Table 7).

We apply optical flow warping to the masks and measure IoU between frames in all VP-UCF101 and VP-HMDB51 clips, and found that the mask refined by SAM2 is more temporally coherent than the coarse Privacy Masker predictions. Under fast motion, the IoU decreases by only 0.094 for SAM2 versus 0.213 for coarse masks.

Temporal mask consistency before and after SAM2 refinement. Higher flow-warped IoU is better.

DatasetCoarse MaskSAM2-Refined Mask
VP-UCF1010.6570.885
VP-HMDB510.6640.865

Edited video flickering

We evaluate temporal smoothness on the full test splits using the VBench[R1] temporal flickering score, computed from the pixel similarity between consecutive video frames (higher is better). The results suggest that thanks to the generative step, our anonymized output maintains temporal coherence comparable to the original videos, regardless of whether coarse (without SAM2) or refined (with SAM2) privacy masks are used.

MethodSAM2 RefinementVP-UCF101 ↑VP-HMDB51 ↑
Original video—97.8997.85
Mask2Safe (Generative)✗97.8997.84
Mask2Safe (Generative)✓97.9297.79

Generated appearance consistency (Appendix G, Table 12).

To assess whether video-level anonymization maintains consistent facial appearances across frames, rather than introducing the identity fluctuations associated with independent framewise editing, we compute within-video ArcFace embedding similarity [R2] across matched face detections from six sampled frames per clip:

Within-video facial appearance consistency measured by ArcFace cosine similarity. Higher is better.

DatasetFramewise Unsafe2SafeVideo-Level Mask2Safe
VP-UCF1010.3130.599
VP-HMDB510.3050.578

We provide an additional evaluation of cross-video identity leakage using ArcFace embeddings (Appendix G). Specifically, we measure facial similarity across different videos containing the same person. Ideally, anonymization should make it difficult to recognize that the same person appears in multiple videos. On VP-UCF101 and VP-HMDB51, cross-video facial similarity (lower is better) decreases from 0.769 to 0.223 and from 0.722 to 0.224, respectively, after generative anonymization. These results indicate considerably reduced identity linkability across videos.

Hence, in addition to maintaining a better temporal flickering score, video-level generation produces more consistent facial appearances across frames than framewise editing. Meanwhile, the low similarity to original faces indicates that this temporal consistency does not come at the expense of preserving the original facial identity.

W6. Comparison With Penalty-Driven Anonymization (Related Work, Section 2)

The penalty-driven image anonymization approach of Aslam and Nasrollahi (CVPRW 2025), together with the subsequent video token-pruning approach of Aslam et al. (CVPRW 2026), is cited and discussed in Section 2 of our original submission. These methods investigate privacy-utility preservation through penalty-based learning or token selection. Mask2Safe explores another direction based on learning privacy-sensitive spatial localization from unsafe-safe image pairs and replacing selected content through generative editing, without requiring downstream action supervision to train the Privacy Masker.

Although the implementation repository for this work is available, the authors have not released pretrained checkpoints or pre-computed anonymized outputs that we could use to reliably reproduce their results under our evaluation protocol. Instead, we focused on baselines with published model weights (TED-SPAD) and reproducible pipelines (Unsafe2Safe), as described in Sec. 5.1 (Experimental Setup-Evaluation Metrics).

We expanded the related work discussion in Section 2 of the main paper (highlighted in blue) to clarify these methodological distinctions and the scope of the available comparisons.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.