acceptodds
Under review as a conference paper at ICLR 2027

ZeroFace: Unified Privacy-Preserving Framework for Secure Face Verification

Abstract

Face recognition (FR) is widely used in authentication and surveillance applications due to its high accuracy for identity verification and recognition. However, there are potential privacy risks existing within FR systems at different stages, particularly when sensitive biometric data can be accessed without authorization. Existing countermeasures focus on protecting users' privacy by concealing identity information that underlies the facial embedding. However, recent generative privacy attacks are able to reconstruct high-quality and identity-preserving face images from facial embeddings. In this work, we propose a secure and efficient framework, ZeroFace, for embedding protection. Firstly, a compression network is proposed to reduce the embedding dimensionality by mapping it into low-dimensional representations, while maintaining nearly lossless recognition accuracy via learning the multi-scale identity information. Furthermore, identity-specific transformation is proposed to further obfuscate the compressed embedding. Lastly, a secure matching protocol is designed to enhance embedding security by secretly computing cosine distance without sharing the user's compressed embedding with the database holder. Extensive experiments show ZeroFace largely maintains recognition accuracy and enhances the embedding robustness compared to the state-of-the-art face template protection algorithms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
face embedding, privacy attack, privacy protection, face recognition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ad4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ZeroFace, a privacy-preserving framework for face verification that combines three components: Adaptive Matryoshka Compression (AMC) for reducing a 512-dimensional face embedding to a compact representation, IS-DOT for identity-specific protection through bounded DCPE perturbation and orthogonal transformation, and Secure Cosine Distance Computation (SCDC) for matching secret-shared templates using Beaver triples.

The method is evaluated on standard face-verification benchmarks including LFW, CFP-FP, AgeDB, CALFW, CPLFW, IJB-B, and IJB-C, as well as against FEM embedding-inversion and MAP2V score-based reconstruction attacks. The reported results show that the 128-dimensional representation retains recognition performance close to the underlying ArcFace model and substantially reduces reconstruction attack success compared with several template-protection baselines. Additional analyses consider soft-biometric attribute leakage, unlinkability, computational/storage cost, compression dimensionality, IS-DOT components, and AMC adaptation to other recognition backbones.

Critical strengths
  1. The manuscript evaluates both embedding-inversion and score-based reconstruction using recent FEM and MAP2V attacks rather than relying only on reconstruction error or conventional template-security measures. Table 2 shows substantial reductions in attack success rates relative to MLP-Hash, PolyProtect, and SlerpFace under the evaluated settings. This is consequential because resistance to modern generative reconstruction is central to the paper's motivation.

  2. Despite reducing the ArcFace embedding from 512 to 128 dimensions, the proposed representation remains close to the unprotected model across LFW, CFP-FP, AgeDB, CALFW, CPLFW, IJB-B, and IJB-C. The paper also includes meaningful analyses of dimensionality, attribute leakage, unlinkability, IS-DOT components, and adaptation to several other FR/PPFR backbones, giving a more complete picture than a single utility-versus-privacy table.

Critical weaknesses
  1. AMC combines linear projections of nested prefixes using global learned scalar weights. Consequently, the complete mapping is itself an affine transformation. The paper compares AMC with PCA, an MLP, and Naive MRL in Table 5, but does not compare with the most important control: a single Linear(512,128) layer trained with exactly the same VGGFace2 data, ArcFace supervision, optimizer, and frozen recognition backbone. Without this comparison, it is unclear whether the reported benefit comes from the proposed Matryoshka structure or simply from supervised dimensionality reduction. This is central to the paper's claimed methodological contribution.

  2. In Table 2, MLP-Hash, PolyProtect, and SlerpFace are attacked through their exposed protected representation or score, while the full ZeroFace setting assumes that a compromised party observes only one additive secret share. A single random share is intentionally insufficient to reconstruct the protected embedding, so the near-zero ASR is not directly comparable to attacks on complete protected templates. Furthermore, Section 3.4 states that basic SCDC can reconstruct the similarity score and that hiding the score requires the optional TEE. It is therefore unclear whether the 0% score-based ASR in Table 2 should be attributed to SCDC itself or to the SCDC+TEE decision-only configuration. The paper should separate representation robustness, secret-sharing confidentiality, and score suppression.

Questions for authors
  1. Since AMC consists of linear branches combined using global learned scalar weights, the complete mapping can be collapsed into a single affine 512-to-128 transformation. What recognition and FEM-ASR results are obtained using a single Linear(512,128) layer trained with exactly the same VGGFace2 data, ArcFace supervision, optimizer, epochs, and frozen IR-50 backbone? This comparison would materially affect my assessment of AMC's novelty.

  2. Can the authors separate Table 2 into results for: (i) AMC+IS-DOT with the complete protected embedding exposed, (ii) SCDC with only one additive share exposed, (iii) basic SCDC with the exact similarity score released, and (iv) SCDC+TEE with decision-only output? In particular, was the reported 0% score-based ASR for “Ours” obtained using the optional TEE?

  3. Where are the identity-specific parameters (Kudcpe,su,ξu)(K_u^{dcpe}, s_u, \xi_u) stored, and how does a legitimate client obtain the parameters corresponding to claimed identity uu? Under the client-compromise scenario in Figure 1, does the adversary obtain these parameters, and what protection remains if both the protected embedding and identity-specific key are compromised?

  4. The irreversibility analysis reports 0% SAR using a numerical solver with only ten iterations. Have the authors evaluated stronger known-operator attacks such as pseudoinverse recovery, longer embedding optimization, or an inversion network explicitly trained against AMC?

  5. Has identity overlap between the VGGFace2 dataset used for AMC training and the verification datasets been checked or removed? Because AMC is trained with identity supervision, confirmation of identity-disjoint evaluation would strengthen the generalization claim.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer ad4V

Thanks for the constructive and valuable comments.

Clarification of Table 2, in our paper, we excluded TEE of SCDC in all the experiments.

W1&Q1 AMC:

We agree that AMC can be potentially approximated to an affine transformation. As the reviewer suggested, we include a baseline model using a Single Linear layer (Single Linear(512,128)) trained in the same way. AMC achieves higher verification accuracy than the Single Linear layer and demonstrates greater robustness to the FEM attack on M.Face, with an ASR of 47.9% compared with 75.5%. This demonstrates the benefits of training the model at multiple scales using AMC.

MethodLFWCFP-FPAgeDB
Single Linear99.5797.1496.38
AMC99.7397.4197.68
MethodFEM-ASR
Single Linear75.5
AMC47.9

W2&Q2 Table 2:

TEE is not included in our two tested settings in Table 2. Therefore, the 0% score-based ASR for “Ours” obtained without using TEE. We would like to clarify the requirements for both TI attacks considered in our work. The inversion-based attack utilizes embeddings, while the score-based attack utilizes the cosine scores for image reconstruction. As the reviewer suggested, we report different settings as below. AMC+IS-DOT is already included in Table 2.

MethodInversion-based ASR ↓ — E.FaceInversion-based ASR ↓ — M.FaceScore-based ASR ↓ — E.FaceScore-based ASR ↓ — M.Face
Unprotected92.699.387.398.1
AMC+IS-DOT0.40.50.642.3
SCDC (one additive share)0.00.40.00.0
SCDC (exact similarity score)0.00.10.037.1
SCDC+TEE (decision-only)0.00.00.012.8

Since the score-based attack MAP2V relies on zeroth-order gradient estimation, it requires continuous cosine similarity scores to estimate gradients and iteratively optimize the StyleGAN latent codes. In the decision-only output setting, accessing only the binary output can greatly disrupt the attack's effectiveness.

Q3 IS-DOT Key:

The identity-specific parameters are generated during enrollment and stored as the local record on the client side, e.g., (ID1, IS-DOT key), (ID2, IS-DOT key). During 1:1 verification, the legitimate client retrieves the IS-DOT key from the local record for the claimed identity. In our experiments, the adversary's goal is to reconstruct a complete face image from the leaked protected embedding.

In the client-compromise scenario in Figure 1, we assume the adversary can submit images to the system and obtain corresponding protected embeddings. The adversary has no knowledge of the structure of AMC and IS-DOT. In the inversion-based attack, we assume the adversary has access to IS-DOT keys of identities in the training dataset (FFHQ). Although the adversary can train an FEM model, it cannot generalize well to another dataset (CelebA-HQ) due to unseen identities, see performance in Table 2.

Q4 Irreversibility:

In the irreversibility experiment, we randomly initialized ten solver runs, each with its own internal iterations. We will revise this description. The current inversion-based attack (FEM) presented in our work is a learned inversion network. We trained FEM models on compressed embeddings produced by AMC with different levels of compression, see results in Figure 5. AMC can reduce ASR from 99.1% (attacked based on the raw embedding) to 47.9% on 128-d compressed embedding. Thus, ZeroFace consists of AMC, IS-DOT and SCDC for enhanced protection.

Q5 Identity-disjoint Evaluation:

We thanks reviewer’s suggestion. We filtered out overlap identities between LFW and VGGFace2 when training AMC. In addition, we also trained the AMC on MS1M dataset using the same training configuration.

MethodLFWCFP-FPAgeDBCALFWCPLFW
ArcFace99.7398.0097.8795.9292.50
AMC (VGGFace2)99.7397.4197.6895.9592.02
AMC (MS1M)99.6096.7497.1795.6591.33

Thank you. The added baseline and clarifications partly address my concerns.

However, several issues remain. First, AMC is exactly affine under the described structure, so please distinguish its training benefits from its final mapping. Second, the 0% score-based ASR applies to the single-share setting. With exact scores released, ASR reaches 37.1% on M.Face. Please clarify the attacker’s access in each setting and revise the security claims accordingly.

Third, the response does not establish protection when the target identity’s key and protected template are both compromised and the operators are known. Please evaluate this setting and stronger known-operator recovery attacks. Finally, please confirm identity-overlap checks for all evaluation datasets and state whether LFW filtering applied to the original experiments.

Effeteness of AMC

We agree that the final mapping from AMC is affine before L2 normalization. Inspired by MRL’s hierarchical representations, AMC uses a multi-granularity training parameterization. The effectiveness of this training parameterization can be observed from the previous Single Linear layer experiment. AgeDB achieves improved verification accuracy of 96.38% to 97.68% and M.Face achieves reduced FEM-ASR of 75.5% to 47.9%. We will distinguish the benefits of the training scheme.

Attacker’s Access

We thank reviewer’s detailed comments on attacker’s access in TI attack experiments. To distinguish the different protection settings, we explicit restrict the information that accessed by the attacker.

Method / settingAttacker accessInversion-based ↓ E.FaceInversion-based ↓ M.FaceScore-/decision-based ↓ E.FaceScore-/decision-based ↓ M.Face
UnprotectedOriginal embedding + exact scores92.699.387.398.1
AMC+IS-DOTProtected embedding + exact scores0.40.50.642.3
SCDC (single additive share)One additive share only0.00.4N/AN/A
SCDC (exact score)Exact scores onlyN/AN/A0.037.1
SCDC+TEE (decision only)Binary accept/reject decisions onlyN/AN/A0.012.8

N/A indicates that the evaluated attack is not applicable as its required information is unavailable.

-Unprotected: The attacker has access to both the original embedding and similarity scores. This allows both the inversion-based and score-based attacks.

-AMC+IS-DOT: The attacker has access to both the original embedding and similarity scores. This allows both the inversion-based and score-based attacks.

-SCDC (Single additive share): The attacker has access to only a single additive share. This allows the inversion-based attack using this share.

-SCDC (Exact score): The attacker has access to similarity scores. This allows the score-based attack.

-SCDC+TEE (Decision only): The attacker has access to only the accept/reject decision. This allows the score-based attack by replacing the score feedback as binary match in the MAP2V attack pipeline.

Overall, SCDC itself remains vulnerable to score-based attacks when attacker has access to exact similarity scores, with an ASR of 37.1% on M.Face.

Strong Known-operator Attack

We thank reviewer’s suggestion for evaluating on the strong known-operator attack. We implement a pseudoinverse attack (based on Moore–Penrose pseudoinverse) to evaluate AMC+IS-DOT under different attack settings regarding to knowledge of AMC weight and IS-DOT key on AgeDB evaluation dataset. We consider four combinations of attack exposure. In the unknown-weight setting, a randomly initialized AMC surrogate is used. We assume attacker has white-box access to structure/algorithms of AMC and IS-DOT.

AMC weightIS-DOT keySAR
knownknown98%
knownunknown0%
unknownknown0%
unknownunknown0%

These results indicate that exposing model structure, model weight and key at the same time can potentially break embedding protection. Keeping the IS-DOT key secret reduces the SAR to 0%, even when the AMC weights are known, demonstrating its contribution against this pseudoinverse recovery attack.

Overlap Identities

In our work, we used VGGFace2 training dataset and evaluation datasets provided by ArcFace paper [1]. All overlap identities between training dataset and evaluation datasets are removed. All our original experiments are conducted under this setting.

[1] 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. 2019.

We thank again for reviewer’s careful reading of our paper. We would be glad for any further discussion.

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

The paper proposes a method for protecting facial recognition features, which consists of three sub-modules: AMC, used to compress original features into low-dimensional representations; IS-DOT, employed for feature transformation by introducing random perturbations; and SCDC, designed for secure cosine distance computation. The experimental results demonstrates good performance in maintaining face recognition accuracy, with the effectiveness of feature protection verified through attack experiments.

Critical strengths

None

Critical weaknesses

The method proposed in this paper is a combination of existing approaches. (1) The concept of feature compression has already been proven effective in prior work—such as SlerpFace—and the AMC structure employed in the paper is also quite common. Furthermore, the comparisons made during the training of this module are somewhat unfair; as noted in the training description, the recognition model’s backbone was trained on the MS1M dataset, while the AMC structure utilized VGGFace2. This means that more training data was used, which raises questions regarding the true recognition performance of the proposed method. (2) The IS-DOT mechanism adopted in the paper appears to require the storage of a unique key for each user. However, the paper fails to provide a security analysis for scenarios where these keys might be compromised. (3) The introduction of SCDC alters the established paradigm in this field; if multi-party computation (MPC) can be employed, the preceding two steps may not even be necessary. (4)The method is still limited to 1v1 scenarios, which significantly restricts its scope of application.

Questions for authors

My main questions are: (1) How is the key Θu generated for each user, and how is security guaranteed when this key is leaked (similar to feature leakage)? (2) Does the storage comparison given in the paper include the storage of the aforementioned key?

1 comment
AuthorsOpenReview ↗

Responses to Reviewer socm

Thanks for the constructive and valuable comments.

W1 Novelty and Additional Supervision:

Regarding the feature compression, SlerpFace produces the output embedding of size 7×7×167\times7\times16, which is equal 784-d embedding after flattening, compared with ours (128-d embedding). It even increases the feature dimension compared with the FR model's original output 512-d embedding. Besides, SlerpFace requires modifying the structure of the FR model and retraining the entire model. Our method only requires training AMC while the FR model remains frozen. Our contribution is a embedding-protection and secure-verification framework that can be adapted without retraining the recognition backbone.

In addition, we also trained the AMC on MS1M dataset using the same training configuration.

MethodLFWCFP-FPAgeDBCALFWCPLFW
ArcFace99.7398.0097.8795.9292.50
AMC (VGGFace2)99.7397.4197.6895.9592.02
AMC (MS1M)99.6096.7497.1795.6591.33

W2&Q1,2 Identity Keys:

The identity-specific parameters are generated during enrollment and stored as the local record on the client side, e.g., (ID1, IS-DOT key), (ID2, IS-DOT key). For the details of IS-DOT key generation, please refer to Appendix C. During 1:1 verification, the legitimate client retrieves the IS-DOT key from the local record for the claimed identity. Even if the key is leaked, an adversary cannot reconstruct the original embedding without knowledge of the structure of IS-DOT and AMC.

In the inversion-based attack, we assume the adversary has access to IS-DOT keys of identities in the training dataset (FFHQ). Although the adversary can train an FEM model, it cannot generalize well to another dataset (CelebA-HQ) due to unseen identities, see performance in Table 2.

In Table 4, we didn’t include the key storage when reporting the storage cost. Specifically, the IS-DOT key consists of a PRF key, secret scaling factor, and secret seed for generating an orthogonal matrix. They are 32 bytes, 8 bytes, and 32 bytes, respectively. Therefore, it takes 72 bytes for each identity.

W3 AMC and IS-DOT with MPC:

MPC can indeed operate directly on the original embeddings. In our work, AMC is proposed to improve the efficiency (reduce database storage (Table 4), improve robustness against TI attacks (Table 2, 3 and Figure 5), and reduce computational cost in SCDC) of secure verification. Embeddings are compressed from 512 dimensions to 128 dimensions to reduce the number of element-wise secure multiplications, Beaver-triple elements and communication payloads in SCDC by 75% while keeping verification accuracy high. Moreover, the SCDC-only or MPC-only setting can expose the original embedding once both parties are compromised. Therefore, AMC and IS-DOT are necessary steps.

W4 Verification and Identification:

We thank reviewer's suggestion. Our main evaluated task is 1:1 verification, as stated in the paper title. We will narrow application claims accordingly. We acknowledge that the current ZeroFace is evaluated on 1:1 verification. To investigate the applicability in 1:N identification, we conducted an additional closed-set 1:N identification experiment on LFW using 500 identities. We use 500 gallery images (one image per identity) and 1978 probe images.

The results are reported below.

MethodRank-1 ↑Rank-5 ↑Rank-10 ↑
Original IR-50 (512-d)98.2898.8498.99
MRL (128-d)84.1392.7794.94
PCA (128-d)97.5298.4398.89
MLP (128-d)90.9096.4197.57
Ours (128-d)97.6798.6998.99
Ours (64-d)94.0397.7298.38

Ours (128-d) achieves 97.67% Rank-1 accuracy, only 0.61% below the raw performance. These results demonstrate that ZeroFace preserves identification utility in 1:N evaluation.

Reviewer_B2NZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work aims to defend against reconstruction attacks on leaked facial embeddings in face recognition systems while preserving verification accuracy. The proposed approach combines dimensionality reduction with perturbation to randomise the compressed representations. It also incorporates multiparty computation to enable verification without directly sharing the exact embeddings.

Strengths • The threat model is clearly defined. • The approach combines embedding compression, perturbation, and multiparty computation. • The evaluation demonstrates good face verification accuracy and considers multiple attacks.

Weaknesses • The three components are not sufficiently integrated. Although each contributes to the approach, the central technical contribution is unclear. • Further evaluation is needed to substantiate the security claims and explain the observed results.

Detailed Comments

  1. The evaluation does not include privacy-preserving face recognition (PPFR) methods that directly protect input facial images. Recent work in this direction can resist diffusion-based reconstruction attacks. A comparison with these methods is needed. The paper should also compare embedding-based reconstruction attacks with diffusion-based reconstruction attacks to better position its contribution within the existing literature.
  2. The related work should discuss other approaches that share only partial information, such as model parameters or selected parts of a model, and explain how the proposed approach differs from them.
  3. The approach combines three components corresponding to three adversarial capabilities. However, the choice of these capabilities appears ad hoc. The paper should justify why they need to be addressed together, explain how the components form a coherent defence, and clarify why other attacks are outside the scope.
  4. Why must the perturbation be bounded by (s_u\beta/4)? If (\beta) is too small, the perturbation may provide insufficient protection. How should (\beta) be selected? Furthermore, given the bounded range of (\lambda), if sus_u is drawn from a sufficiently small or predictable set, an adversary may enumerate candidate original embeddings from the transformed representation. The paper should analyse whether embedding constraints or auxiliary information could enable the adversary to identify the correct candidate.
  5. The SCDC protocol requires a formal security analysis. Without such an analysis, it is difficult to assess whether the protocol provides the claimed security guarantees.
  6. The evaluation shows that IS-DOT is effective in one setting but not another, yet no explanation is provided. The paper should investigate the reasons for this difference and discuss its implications.
  7. The paper categorises existing methods as crypto-based, hash-based, or transform-based. The proposed approach appears to combine crypto-based and transform-based techniques, and its runtime falls between those of the two categories. The paper should explain this trade-off more thoroughly, including whether the additional protection justifies the overhead relative to transform-based methods.

Minor Comments • Page 1: The paragraph beginning “Existing BTP methods can be broadly categorized into three groups…” moves directly to transform-based methods. Please also briefly discuss crypto-based and hash-based methods. • Page 7: SCDS matching takes 2.8 ms. Please contextualise this cost in terms of the latency and throughput requirements of practical applications. • Page 8: Please justify the choice of 10 iterations and clarify whether the results are sensitive to this choice. • Page 8: Please provide the full name of LFW at its first occurrence.

Critical strengths

• The threat model is clearly defined. • The approach combines embedding compression, perturbation, and multiparty computation. • The evaluation demonstrates good face verification accuracy and considers multiple attacks.

Critical weaknesses

• The three components are not sufficiently integrated. Although each contributes to the approach, the central technical contribution is unclear. • Further evaluation is needed to substantiate the security claims and explain the observed results.

Questions for authors

See Detailed comments 1-7.

3 comments
AuthorsOpenReview ↗

Responses to Reviewer B2NZ

Thanks for the constructive and valuable comments.

W1: We would like to clarify that our work mainly focuses on face embedding protection, as stated in the abstract. PPFR is fundamentally different method which focuses on face image protection. Moreover, PPFR has already been studied in previous work such as MAP2V and FEM, and it is vulnerable to GAN-based and diffusion-based reconstruction attacks.

W2: We will discuss related works such as federated learning and split inference. Federated approaches alter what information is shared during collaborative learning, while split inference refers to distributing the model and intermediate representations. ZeroFace works with outputs of FR backbone (embeddings). There are differences between the types of information being protected, trust assumptions, and information revealed.

W3: As mentioned in the abstract, introduction, and threat model sections in our paper, we mainly focus on reducing risks against recent GAN-based and diffusion-based attacks. AMC produces a compact identity-discriminative representation to reduce the identity information that can be exploited for facial reconstruction, in Figure 5. Moreover, due to the embedding compression of AMC, it also reduces the computational cost required by SCDC. Since AMC provides limited protection when adversary has extensive query access to the target FR model. We introduce IS-DOT to add revocable and ID-specific protection so that the adversary cannot reconstruct the face image of an unseen identity even after training the FEM model on specific identities. Besides, IS-DOT allows us to renew the protection by updating the key. Lastly, SCDC enables face matching without revealing the other party's complete protected embedding. Considering a scenario where both parties are compromised in SCDC, the embedding can be leaked. Therefore, AMC and IS-DOT are also necessary steps before SCDC.

W4: DCPE used in our IS-DOT is a distance-comparison preserving scheme. The main characteristic of DCPE is that it introduces perturbation to embeddings while retaining the relative distance. This property is crucial for face verification, as we aim to protect face embeddings while maintaining their utility. The perturbation bound is related to DCPE’s security.

As stated in the DCPE paper, the choice of β\beta depends on the dataset and application. We reported the verification accuracy on different settings, see in Appendix G, Figure 9. Overall, it should be selected according to utility preservation. Based on our observation, β\beta within [0,1] is a good practice.

Considering DCPE alone, a correct scale guess yields an approximate embedding with L2 error at most β/4\beta/4. This auxiliary information may help identify an approximated embedding. However, IS-DOT additionally applies an ID-specific orthogonal transformation, so such auxiliary information cannot further assist in original embedding reconstruction.

W5: The core primitives behind SCDC are described in Appendix A.2. The details of the additively shared, trusted preprocessing, and Beaver triple multiplication protocols used are given in Appendix D (Equations 38-46). These are common MPC building blocks with security proofs already established in previous work Beaver [1] and CrypTen. We will describe how these security guarantees apply to the SCDC protocol in the revision.

[1] Donald Beaver. Efficient multiparty protocols using circuit randomization. In Annual international cryptology conference, pp. 420–432. Springer, 1991.

W6: Regarding the different effects of IS-DOT under MobileFace (M.Face) and ElasticFace (E.Face) evaluators, we have mentioned the difference between these two FR evaluators, in Section 4.3. M.Face is relatively easy to be fooled due to its limited distinguishability. Thus, we observe high ASRs on M.Face, in Table 6.

Regarding the different effects of IS-DOT under inversion-based and score-based attacks, IS-DOT alone does not change too much of the similarity score between two matched protected embeddings compared with the similarity score directly calculated from original embeddings. Therefore, IS-DOT provide limited defense against score-based attack than inversion-based attack.

W7: Our method combines transform-based protection with MPC-based matching to limit disclosure among parties under the assumptions. Adding SCDC reduces score-based ASR from 42.3% to 0% as evaluated by M.Face, providing empirical support for the additional cost. Overall, our method achieves better privacy protection and accuracy preservation than previous methods. We will expand the discussion of this trade-off in the revision.

Minors: We will introduce more about crypto-based and hash-based BTP methods in the revision. The latency also depends on feature extraction and communication. “10 iterations” indicate to 10 random restarts of numerical solver and each run has internal iterations. LFW: Labeled Faces in the Wild.

AuthorsOpenReview ↗

Security Analysis of SCDC

We added a security analysis of SCDC (highlighted in blue) in Appendix D.

Thank you again for your thoughtful comments and valuable suggestions on our paper. We would be happy to address any further questions during the discussion.

Best regards, the authors

AuthorsOpenReview ↗

Comparison with PPFRs

Dear reviewer B2NZ,

We thank your valuable feedback on our paper.

To further answer the question regarding comparison with PPFRs, we conducted experiments by training FEM-MLP models (inversion-based attack, it reconstructs face images from embeddings via the diffusion model and can be seen as an embedding-based reconstruction attack) against PPFRs on FFHQ and evaluated on CelebA-HQ. Our method focuses on protecting face embeddings while PPFRs focus on protecting face images.

As shown in Table below, our method achieves the strongest robustness against inversion-based attack compared with PPFRs. Moreover, these PPFRs require retraining/fine-tuning the FR backbone to retain recognition accuracy, which limits the transferability compared with our method.

Protection MethodVenueASR(%) ↓ E.FASR(%) ↓ M.F
PartialFace [1]ICCV202388.598.7
MinusFace [2]CVPR202468.594.4
LDP-Slicing [3]CVPR202644.267.1
Ours (AMC+IS-DOT)0.40.5

[1] Mi, Yuxi, et al. "Privacy-preserving face recognition using random frequency components." 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023.

[2] Mi, Yuxi, et al. "Privacy-preserving face recognition using trainable feature subtraction." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

[3] Yuanming Cao, Chengqi Li, Wenbo He. “LDP-Slicing: Local Differential Privacy for Images via Randomized Bit-Plane Slicing.” CVPR, 2026.

We hope the added experiment can address your concerns regarding PPFRs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.