acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Identity Targets for Face De-Identification: Attribute Compatibility Matters

Abstract

Face de-identification aims to suppress identity information while preserving the utility of facial images for downstream tasks. Effective de-identification must therefore balance privacy with the preservation of non-identity information, visual fidelity, and realism. To this end, modern de-identification methods increasingly rely on generative models to transform the source image towards a desired target appearance. Here, the target appearance is typically defined by a face-recognition (FR) embedding that is sufficiently different from the source embedding in order to reduce the likelihood of re-identification. However, selecting targets mainly based on their separation from the source ignores the fact that FR embeddings also encode non-identity information, such as age, gender, ethnicity, and pose. As a result, unsuitable targets can induce unwanted attribute changes and reduce the utility and fidelity of de-identified images. In this paper, we introduce HIDE (Hyperspherical Identity Target Editing), a novel flow-matching framework for face de-identification that explicitly accounts for this problem by first constructing a target that is sufficiently separated from the source identity and then refining it to improve compatibility with source attributes. The resulting target is then used to condition HIDE's flow-matching generator that transforms the source image toward an anonymous identity while retaining its attributes. To strengthen identity conditioning, we introduce an online inverse training branch that reconstructs the source image from the generated de-identified sample, using the source as the target for both the flow-matching and the identity objective. We evaluate HIDE in comprehensive experiments against nine recent state-of-the-art face de-identification methods on three diverse face datasets, i.e., LFW, CelebA-HQ, and WFLW. Our results demonstrate that HIDE outperforms existing methods in terms of the privacy–utility trade-off, ensuring strong identity removal while preserving facial attributes and visual fidelity across challenging variations in pose, expression, age, and occlusion. Source code will be made available.

Primary area:
generative models
Keywords:
Face de-identification, Privacy protection, Flow matching, Attribute preservation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Y5HQOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript introduces HIDE, a flow-matching framework for face de-identification. The authors' main argument is that existing methods usually bolt auxiliary constraints onto the generator to balance identity removal and attribute preservation, but pay little attention to whether the target identity is actually compatible with the source face's non-identity attributes; as a result, the training objective itself can be internally inconsistent. HIDE moves this issue to the target-construction stage. It first creates an FR anchor orthogonal to the source identity, then uses a lightweight target editor to edit that anchor on the hypersphere into a target compatible with the source attributes, and finally generates the de-identified image with target-guided conditional flow matching.

Critical strengths

1.Clear and meaningful conceptual contribution: This paper proposes the concept that “source-attribute compatibility” should be treated as a first-class requirement for target construction, rather than merely adding auxiliary losses on the generator side, directly addressing issues such as attribute drift, target appearance contamination, and mixed-identity ghosting in existing target-attraction/source-repulsion methods. 2.Solid experiments: The proposed method evaluates privacy on three datasets, nine baselines, and five FR models, and covers utility metrics such as FID, pose, gaze, age, gender, race, and expression. The ablation and target-space analyses also support the core claims well.

Critical weaknesses

None

Questions for authors
  1. The paper sets theta_max=70 but does not justify this choice. How was this value determined? It is recommended to provide additional explanation.
  2. In Table 1, the LFW ASR of DiffPrivacy and RP is significantly lower than that of other methods. Were these two baseline sufficiently tuned to achieve their best performance?
  3. In Table 1, different methods operate at very different privacy operating points, while utility metrics are computed under their respective default configurations. Directly comparing single-point utility may fail to disentangle “differences in privacy strength” from “differences in utility preservation capability.” The paper claims that HIDE achieves a better privacy–utility trade-off, but it does not adjust the baselines to the same privacy level before comparing utility. Therefore, it is recommended to add comparisons at matched privacy levels or report Pareto frontiers.
  4. Table 2 has already validated the effectiveness of online inverse training through ablation experiments, but the main text discusses these results only briefly. Since online inverse training is listed as one of the three main contributions, it is recommended to expand the main text to explain its specific gains, mechanism.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Y5HQ (Part 1/2)

We thank the reviewer for the thoughtful evaluation and encouraging feedback. We appreciate their recognition of our conceptual contribution in treating source-attribute compatibility as a fundamental requirement for identity target construction, as well as their positive assessment of our experimental evaluation and supporting ablation studies. Below, we address each question and provide the requested clarifications.

Q1. Choice of theta_max

Thank you for the question. The parameter θ_max controls the maximum angular displacement of the edited target from its initial anchor. Its purpose is to prevent excessive changes to the randomly sampled identity anchor while allowing sufficient flexibility to adapt it to the source facial attributes. We set θ_max = 70° as a relatively permissive upper bound, while the actual editing angle is learned and constrained by the identity and attribute objectives (Sec. 3.1).

It is important to note that the editor does not typically approach this bound. As shown in Sec. 4.1, the mean cosine similarity between the edited targets and their initial anchors is 0.93, corresponding to an angular separation of approximately 21°. This indicates that the learned edits are substantially smaller than the permitted maximum, suggesting that the method does not rely on large changes to the initial anchors. A discussion of how we selected θmax⁡\theta_{\max} is provided in Fig. 11 (Appendix A.6).

Q2. Baseline Configuration

We thank the reviewer for pointing this out. Among the evaluated implementations, RP is the only method supporting different input resolutions. We initially evaluated it at 512x512 and have now added results at 1024x1024. This substantially improves performance, increasing LFW ASR at FAR 10^-2 from 13.4% to 96.2%, as the method can better exploit SD-XL prior knowledge. We now report results with both resolutions in the revised paper.

The original DiffPrivacy paper reports approximately 20 minutes of MSI optimization per image on an NVIDIA GeForce RTX 3090. For approximately 50,000 images, this would require an estimated 16,700 GPU-hours, excluding sampling. We therefore used the authors' supplied example MSI checkpoints from CelebA-HQ dataset without per-image optimization, retaining embedding scheduling and ArcFace-based identity repulsion during DDIM sampling. We clarified this evaluation in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer Y5HQ (Part 2/2)

Q3. Privacy-Utility Trade-off

We thank the reviewer for highlighting the importance of comparing methods at comparable privacy levels. We agree that evaluating methods only at their default configurations does not fully characterize their privacy–utility trade-offs.

To address this concern, we conducted additional experiments in which we varied the anonymization strength of HIDE and competing methods that support this functionality during inference. The results, presented in the new Fig. 10 (Appendix A.5), allow us to compare utility preservation across different privacy operating points. They show that competing methods generally exhibit more pronounced attribute degradation as privacy increases, whereas HIDE maintains high utility even at stronger privacy levels. This supports our conclusion that HIDE achieves a more favorable privacy–utility trade-off, rather than merely benefiting from a particular operating point.

It is important to note that Table 1 uses the configurations recommended by the respective authors, which we consider the fairest basis for the primary comparison, since modifying these settings may lead to suboptimal performance. At these configurations, HIDE achieves both stronger privacy and superior overall utility preservation. Matching HIDE's privacy level would generally require competing methods to increase their anonymization strength, potentially leading to further utility degradation, as illustrated in Fig. 8. We would therefore like to retain Table 1 as the main comparison and provide the additional trade-off analysis in Appendix A.5.

Q4. Online Inverse Training

We thank the reviewer for recognizing the effectiveness of online inverse training. We have expanded the discussion in Sec. 3.3 and Sec. 4.2 to better explain its mechanism and contribution. Specifically, the inverse branch reconstructs the source image from an online-generated de-identified image while conditioning on the original source identity. This provides aligned reconstruction and identity supervision, encouraging the generator to follow the specified identity condition more effectively. As demonstrated in Table 2 and Fig. 6, inverse training primarily improves identity suppression, while introducing only minor changes in attribute preservation.

AuthorsOpenReview ↗

Summary of Our Response to Reviewer Y5HQ

Thank you again for the encouraging evaluation and constructive suggestions. We clarified the selection and practical effect of theta_max, provided the stronger RP configuration at 1024x1024, and added matched privacy–utility analyses across operating points. We also expanded the explanation of online inverse training and its role in strengthening identity suppression. We hope these clarifications and results address the requested points, and we would appreciate the reviewer’s reassessment of the overall recommendation in light of the rebuttal and revised manuscript.

AuthorsOpenReview ↗

Kind Follow-up on Our Response

Dear Reviewer Y5HQ,

Thank you again for your encouraging assessment and constructive suggestions.

As the discussion period is approaching its end, we would like to kindly follow up on our rebuttal. We hope our responses and the additional results have adequately addressed your questions. If any questions or concerns remain, we would greatly appreciate your feedback and would be happy to provide further clarification before the discussion period concludes.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_bQ7XOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HIDE, a flow-matching framework with hyperspherical identity-target editing, to address incompatibility between identity targets and source-image attributes in face de-identification. It samples from a PCA model of training-set face-recognition embeddings and projects the sample to obtain an anchor orthogonal to the source identity. A source-conditioned spherical editor then improves attribute compatibility under identity-separation objectives. The edited target guides generation, while an online inverse training branch reconstructs the source image using the generated de-identified image as context and the source embedding as the identity condition. Experiments on LFW, CelebA-HQ, and WFLW demonstrate promising de-identification and attribute-preservation performance.

----- after rebuttal --- The response has clarified and addressed most of my concerns. After careful reconsideration, I believe these issues are addressable, and the authors have tried to address them.

Critical strengths
  • The paper addresses attribute compatibility during identity-target construction, rather than relying solely on generator-side attribute supervision. This provides a clear motivation and a plausible explanation for the attribute changes and ghosting caused by unsuitable targets.

  • The ablations in Table 2 are well designed. The editor is replaced with the original orthogonal anchor while generator-side supervision is retained, helping isolate its contribution. Attribute supervision and inverse training are also evaluated separately.

  • Privacy is evaluated using five recognizers distinct from the ArcFace model used during training, while utility is assessed with independent models. The target-space analysis also uses separately trained and frozen predictors. These choices strengthen the experimental evidence.

  • Figure 2 clearly illustrates the three stages of the method. The appendix provides per-recognizer and per-dataset results, along with computational costs for the complete inference pipeline on aligned inputs, facilitating closer assessment of the method.

Critical weaknesses
  • The CelebA-HQ retrieval results in Appendix Table 4 contain an apparent inconsistency. For HIDE, CosFace yields R@1 = 0.02% and R@5 = 0.00%, while CurricularFace yields 0.01% and 0.00%, respectively. Under the stated protocol with identical queries and galleries, a rank-one hit must also count as a rank-five hit. Thus, R@5 should not be lower than R@1. These entries should be checked and corrected, with any implications for the reported results clarified.

  • The evaluation protocol is not fully specified. In particular, the final numbers of retained CelebA-HQ identities, gallery images, and queries are missing, as is the rule for selecting gallery images for each identity. Providing these details would improve reproducibility.

  • Each competing method is evaluated using a single recommended or released configuration, whereas the strength sweep in Appendix A.1 is conducted only for HIDE. Table 1 therefore demonstrates advantages at the tested configurations but does not establish superior utility at matched privacy levels. The conclusions should be qualified accordingly, or supported by comparisons across privacy–utility trade-offs.

  • The interpretation of the utility metrics requires clarification. Under the paired-original evaluation protocol, the reported gender, race, and expression “accuracy” appears to measure prediction agreement between original and de-identified images, rather than accuracy against independent ground-truth labels. The terminology and associated conclusions should reflect this distinction.

  • Some figure descriptions need additional detail. Figure 6 does not specify the recognizer used to compute cosine similarity or the number of samples included in the analysis. Figure 7 also introduces FIQ without defining it. Clarifying these details would make the results easier to interpret.

Questions for authors

see weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer bQ7X (Part 1/2)

We thank the reviewer for the thorough and constructive evaluation. We appreciate their recognition of the motivation behind attribute-compatible identity target construction, the careful design of our ablation studies, and the comprehensive evaluation using independent models. We also thank the reviewer for identifying areas where the experimental reporting and evaluation protocol can be improved. Below, we address each concern individually.

Q1. Retrieval Metric Inconsistency

We thank the reviewer for identifying this error. The R@1 and R@5 values for HIDE were mistakenly swapped in Table 5. We have corrected these entries. The swap has not affected the averages in Table 1 and the conclusions.

Q2. Evaluation Protocol Clarification

We thank the reviewer for highlighting the current description of the evaluation protocol was not complete. For CelebA-HQ, we perform identity-level retrieval experiments and retain identities with at least two images, resulting in 4,516 identities and 28,229 images. For each identity, one original image is randomly selected as the gallery reference, yielding 4,516 gallery images, while the de-identified versions of the remaining 23,713 images serve as queries. To ensure reproducibility, we sort filenames before sampling and evaluate over five fixed random seeds (2026–2030). The same gallery/query splits are used across all methods and FR models, and results are averaged over the five splits. We have included these details in Appendix A.2.2 of the revised manuscript.

Q3. Matched-Privacy Comparison

We thank the reviewer for highlighting the importance of comparing methods at different privacy strengths, particularly those that allow anonymization strength to be adjusted during inference.

To address this point, we conducted additional privacy–utility trade-off experiments across different anonymization strengths for HIDE and competing methods, presented in the new Fig. 10 (Appendix A.5). The results show that increasing anonymization strength generally leads to more pronounced degradation of attribute preservation in competing methods, whereas HIDE maintains high utility even at stronger privacy levels. Consequently, HIDE achieves a more favorable privacy–utility trade-off across the evaluated settings.

It is important to note that for the initial evaluation in Table 1, we used the hyperparameter settings recommended by the authors of the respective methods. We believe this provides the fairest comparison, as these settings are typically selected to achieve the best performance of each method, while modifying them may lead to suboptimal results. At these recommended settings, HIDE achieves both stronger privacy and superior overall utility preservation compared to existing methods. We therefore retain Table 1 as our main comparison, while the additional experiments in Fig. 10 provide a complementary analysis of the privacy–utility trade-offs across different anonymization strengths.

AuthorsOpenReview ↗

Response to Reviewer bQ7X (Part 2/2)

Q4. Utility Metric Interpretation

We thank the reviewer for this observation. Since ground-truth annotations for the considered facial attributes are not available across the evaluation datasets, we assess utility preservation by comparing attribute predictions between de-identified images and their paired originals, as described in Sec. 4 (Evaluation Protocol) and as correctly observed by the reviewer. The objective is to ensure that independent attribute estimation models produce consistent predictions for the original and de-identified images, indicating that the relevant facial attributes have been preserved.

We agree that “prediction agreement” is more precise than “recognition accuracy” for gender, race, and expression, since these metrics measure consistency between predictions rather than correctness against ground-truth labels. We have updated the terminology and corresponding discussion in the revised manuscript accordingly.

Q5. Figure Details

Fig. 6: The cosine-similarity distributions are computed over the entire evaluation datasets, using ArcFace to measure target adherence and five independent FR models to measure source-identity removal. We have clarified this in the figure caption.

Fig. 7: FIQ denotes Face Image Quality and is estimated from the embedding magnitude produced by MagFace. This metric was inadvertently retained in the appendix figure. Since it primarily reflects recognition-oriented rather than perceptual image quality, we have removed it to ensure consistency with the evaluation protocol and Table 1.

AuthorsOpenReview ↗

Summary of Our Response to Reviewer bQ7X

Thank you again for the careful evaluation. We corrected the retrieval-metric swap, fully specified the CelebA-HQ retrieval protocol, added matched privacy–utility analyses, and clarified that the utility measures are prediction agreement rather than accuracy against ground truth. We also clarified the two figure details. We hope these changes address the main concerns, and we would appreciate the reviewer’s reassessment of the overall recommendation in light of the rebuttal and revised manuscript.

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

This paper presents HIDE, a flow-matching face de-identification framework that addresses the mismatch between identity-separated target embeddings and the facial attributes of source images. It first samples an orthogonal identity anchor and refines it through source-conditioned hyperspherical editing. A conditional flow-matching generator then produces de-identified faces, while an online inverse training branch strengthens identity control through cycle-consistent supervision. The authors rely on public datasets LFW, CelebA-HQ, and WFLW to show their performance. The paper is well organized and easy to follow. However, maybe, the idea of the proposed method for choosing anonymous identities while preserving original attributes is not new and has been used before, but with different methods.

Critical strengths
  1. The paper identifies an important limitation of conventional target selection: large identity separation alone does not guarantee compatibility with source facial attributes.
  2. HIDE combines source-conditioned hyperspherical target editing with a flow-matching generator and a training-only inverse branch for stronger identity control.
Critical weaknesses
  1. The novelty on identity control and attribute preservation needs more clarification since similar idea(s) were already used, but with different implementation.
  2. What kind of utility should be preserved for your method? Are there any new considerations on this point besides protection?
  3. As discussed in Abstract‘FR embeddings also encode non-identity information’, how to separate the attribute information from the FR feature based on orthogonal? Although Fig.4 reports improved results, for the final identity, it would always encode some non-identity information by using ArcFace feature.
  4. The target-editor objective in Section 3.1 and the preprocessing details in Section 4 rely on ArcFace embeddings and pseudo-labels for age, pose, gender, and race. The robustness of target compatibility to imperfect or biased attribute estimators is insufficiently examined.
  5. In Section 3.1, Eq. 5 guides target editing using attribute predictions in face-recognition space. However, these proxy predictions may not fully capture fine-grained appearance or other utility-critical properties.
  6. Section 4 reports substantial training costs, while Appendix A.4 provides efficiency measurements. However, the additional training overhead of the inverse branch in Section 3.3 is not clearly quantified.
  7. The proposed method is somewhat complicated and costive for model learning. How to well train the model with multiple loss items?
  8. In Figure 3, the visual results of HIDE seem close to the original faces, which would not good for visual anonymization.
  9. For face recognition, the paper used ArcFace. What will happen if we use more recent FR models for re-identification? How about the robustness?
  10. Can HIDE generate diverse anonymized identities for the same source image while maintaining consistent privacy protection and attribute preservation?
  11. When choosing a different initial identity anchor, will this affect the anonymization performance?
Questions for authors

See weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer TfGG (Part 1/3)

We thank the reviewer for the thoughtful evaluation and constructive feedback. We appreciate their positive assessment of the paper's organization and soundness, as well as their recognition of the importance of attribute-compatible identity selection and the contributions of our paper. Below, we address each concern and question individually.

Q1. Novelty of Identity Control

Thank you for the comment. We clarify that our novelty is not in using FR embeddings for identity control, but in making the target identity embedding compatible with the source attributes before generation of the de-identified image.

FR embeddings encode not only identity but also other facial attributes. Consequently, choosing a target only because of well separation from the source embedding and guiding the generator toward it, as done in prior methods (Sec. 2), can introduce attribute inconsistencies and visual artifacts (Figs. 3 and 5).

Unlike prior approaches, which primarily rely on generator-side constraints for attribute preservation, our HIDE method explicitly adapts the target identity embedding to the source attributes before generation. This fundamental shift makes it possible for HIDE to achieve high identity suppression while retaining utility, yielding state-of-the-art performance in both privacy and utility compared to existing works.

Importantly, our ablation in Table 2 isolates the added value of adapting the target identity itself, by removing target editing while retaining generator-side attribute supervision. Furthermore, the target-space analysis in Sec. 4.1 demonstrates that editing substantially improves attribute compatibility directly in FR space, before image generation.

To the best of our knowledge, explicitly optimizing target-identity embeddings for compatibility with source facial attributes has not been explored in prior face de-identification literature.

Q2. Utility Definition and Considerations

Utility refers to preserving non-identity information required for downstream tasks, including attributes such as pose, gaze, expression, age, gender, and race, alongside visual quality, as evaluated in Table 1. We do not propose a new definition of utility. Rather, our contribution is to improve the preservation of these six facial attributes while achieving strong identity suppression and producing visually convincing, realistic de-identified images.

Q3. Identity-Attribute Disentanglement Clarification

Thank you for the comment. We would like to clarify that orthogonal projection is not intended to separate identity and attribute information within the FR embedding. Instead, it provides an initial target identity embedding that is well separated from the source. Specifically, we start with a randomly sampled embedding and remove its component along the source embedding’s direction. This produces an orthogonal identity anchor with zero cosine similarity to the source embedding, ensuring an identity-separated initialization. However, since FR embeddings encode both identity and non-identity information, orthogonal projection does not control the facial attributes encoded in the resulting anchor, which may therefore be arbitrary. This is stated below Eq. (1) in the main paper.

Once the anchor is defined, we use a target editor that adjusts the anchor to encode facial attributes consistent with those of the source, while maintaining identity separation. Fig 4. shows how the edited identity embeddings encode source attribute information, while random targets do not.

AuthorsOpenReview ↗

Response to Reviewer TfGG (Part 2/3)

Q4. Robustness to Imperfect Attribute Estimators

Thank you for raising this important point. We agree that imperfect or biased attribute estimates may affect target editing and consequently the quality of the generated de-identified images. Since ground-truth annotations for the considered facial attributes are not available across our training datasets, we rely on pseudo-labels obtained from pretrained attribute estimators.

To investigate the sensitivity of HIDE to such inaccuracies and biases, we conducted an additional robustness experiment in which we introduced controlled perturbations to the source embedding before conditioning the target editor. Because the source embedding is the primary source of information used to align the anchor embeddings with the source attributes, such perturbations allow us to gain deeper insight into how inaccurate or distorted source-attribute information affects target editing and the resulting de-identified images, providing insight into the potential consequences of imperfect attribute supervision. The results are presented in Fig. 8 and Table 3 of the updated manuscript, including qualitative examples illustrating how perturbations to the source embedding influence the generated faces. The results show that smaller perturbations have little effect on attribute preservation, whereas larger perturbations lead to increasingly pronounced attribute drift, particularly in age and gender.

Importantly, our existing evaluation also uses independent FR models and attribute estimators (Sec. 4), demonstrating strong identity suppression and state-of-the-art utility preservation without relying on the models used for training. The additional experiment provides further insight into HIDE's sensitivity to inaccuracies in the source-attribute information used for target editing.

Q5. Attribute Predictions and Appearance Preservation

Thank you for the comment. Eq. 5 explicitly optimizes target embeddings to align attributes encoded in FR space, such as age, gender, race, and pose, with those of the source. It is not intended to capture all fine-grained appearance information, as not all such information is necessarily encoded in FR embeddings, as correctly observed by the reviewer. Broader appearance preservation is encouraged through other mechanisms, including source-image conditioning, the flow-matching reconstruction objective, perceptual loss, and image-space attribute supervision.

The effectiveness of the overall approach is demonstrated in Table 1, where HIDE achieves strong preservation of facial attributes across three datasets. Qualitative results in Figs. 3 and 5 further illustrate the preservation of source appearance under challenging conditions, including variations in expression, pose, age, and occlusion. These results show that HIDE effectively preserves a broad range of facial attributes and appearance characteristics, even though the target-editor objective explicitly considers only a subset of the information relevant to appearance preservation.

Q6. Inverse Branch Training Overhead

Thank you for the comment. Training without the inverse branch took approximately 4.6 days, compared with 7.7 days for the full model, corresponding to an additional training overhead of approximately 67%. We clarified these time requirements in Appendix A.4. Importantly, the inverse branch is used only during training and therefore introduces no additional computational overhead at inference.

Q7. Training with Multiple Loss Functions

HIDE combines several loss terms to balance identity suppression, attribute preservation, and image quality. Each serves a specific purpose: the flow-matching loss provides generative supervision, the identity loss enforces anonymization, while perceptual and attribute losses encourage preservation of source appearance. Combining multiple objectives is common in face de-identification methods, and HIDE follows this established practice.

As described in Sec. 4, we did not perform extensive loss-weight optimization. Instead, we followed common practice by assigning higher weights to identity-related objectives and lower weights to attribute-preservation losses. The strong results suggest that extensive tuning is not necessary to achieve competitive performance.

AuthorsOpenReview ↗

Response to Reviewer TfGG (Part 3/3)

Q8. Visual Anonymization

Thank you for the comment. HIDE aims to suppress biometric identity while preserving non-identity facial attributes, such as pose, expression, age, and gaze. Consequently, some visual similarity to the source is expected and desirable for utility preservation, but this does not necessarily imply identity retention. Table 1 demonstrates strong identity removal across five independent FR models distinct from the training model, alongside high utility preservation. We have additionally included average cosine similarities to the source identity in Fig. 3 to provide further quantitative evidence of identity removal alongside the qualitative comparisons.

Q9. Robustness to More Recent FR Models

Thank you for the question regarding more recent FR models. For training, we use identity embeddings extracted with ArcFace (2019), following common practice in face de-identification and facilitating comparison with existing methods. However, to assess robustness against re-identification, we already evaluate HIDE using five independent FR models: AdaFace (2022), CosFace (2018), CurricularFace (2020), MagFace (2021), and UniFace (2023). As shown in Tables 4 and 5 (Appendix A.2), HIDE consistently achieves strong identity suppression across all five independent FR models, demonstrating robust de-identification performance beyond the FR model used for training.

Q10. Identity Diversity

Thank you for the question. Resampling the FR anchor (Sec. 3.1, Eq. 1) enables HIDE to generate different anonymous identities for the same source image. To demonstrate this, we added Fig. 7, showing different de-identified faces obtained using independently sampled anchors.

We further conducted an experiment using 20 independently sampled anchors for each of 100 CelebA-HQ source images. The results in Table 3 show an average cosine similarity of 0.25 between different de-identified versions of the same source, demonstrating identity diversity. At the same time, the average source–output cosine similarity remains low (−0.21), while the attribute-preservation results show that the generated images retain source facial attributes across different anchors.

Q11. Anchor Sensitivity

Thank you for the question. In our main experiments, we independently sample a random identity anchor for each source image, demonstrating strong privacy and utility performance without requiring any specific anchor selection. We have clarified this procedure in Sec. 4.

To further investigate sensitivity to anchor selection, we added an experiment in Sec 4.3 (Table 3), where we generate 20 de-identified images using different randomly sampled anchors for each of 100 CelebA-HQ source images. The results show consistently low source–output identity similarity (mean cosine −0.21, std. 0.09), alongside relatively stable attribute-preservation metrics. These findings indicate that HIDE is not strongly sensitive to the choice of initial identity anchor.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.