acceptodds
Withdrawn from ICLR 2027

FaceLinkGen: A Re-evaluation of Identity Leakage in Privacy-Preserving Face Recognition and Face Anonymization Systems Using Simple Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Privacy-preserving face recognition (PPFR) and face anonymization are designed for different goals, but both must retain some identity-related information to remain useful. We show that this remaining information can be learned by an adaptive attacker. We propose FaceLinkGen, a simple distillation-based attack that trains a face recognition model to map protected inputs back to standard face embeddings. FaceLinkGen applies to keyless PPFR systems and perception-preserving face de-identification (De-ID) systems. For PPFR, the recovered embeddings can be used to regenerate faces that match the original person. Across MinusFace, PartialFace, and DecoyFace, the regenerated faces achieve acceptance rates of 81.0–99.0% on Face++ and 74.9–99.2% on Amazon. For De-ID, FaceLinkGen links protected faces to unprotected images of the same person, reaching Recall@1 values of 48.4–89.6% across the evaluated methods. The attack also remains effective when trained with limited paired data. Our results show that protection against a fixed recognizer or reconstruction model is not enough when identity information remains in the protected representation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Privacy, face, security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4jU7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper re-evaluates identity leakage in privacy-preserving face recognition (PPFR) and perception-preserving face de-identification systems. It introduces FaceLinkGen, a unified feature-distillation attack. The evaluation covers three PPFR systems and proves its effectiveness.

Critical strengths
  1. The result may encourage future PPFR work to use stronger optimization, architecture tuning, and adaptive evaluation before claiming resistance to inversion.
  2. FaceLinkGen expands attack coverage. The pixel reconstruction failure does not imply the absence of identity information. Evaluation with commercial APIs, AdaFace, cross-image retrieval, and soft-biometric prediction also reduces dependence on a single recognition model.
Critical weaknesses
  1. The current evaluation does not fully support realistic human-identity recovery claims. Most experiments use MS1M for training and testing with identity-disjoint splits. This establishes generalization to unseen identities but not necessarily to new image sources, demographics, camera pipelines, or deployment domains.
  2. The novelty and contribution of the proposed method are insufficient. FaceLinkGen essentially fine-tunes a pretrained ArcFace student on paired protected-original examples to reproduce teacher embeddings. Most techniques are widely used in prior works.
Questions for authors
  1. How does FaceLinkGen compare with identity-classification fine-tuning, ArcFace/triplet training, contrastive learning, a frozen encoder plus an MLP adapter, and a linear embedding transformation?
  2. For De-ID, what happens when the cross-image off-diagonal terms or original-domain preservation loss are removed, when α\alpha is varied, or when the teacher and student use different initializations?
  3. Can the method be evaluated on a real, identity-disjoint external dataset, especially for De-ID? A training-on-MS1M and testing-on-another-real-dataset protocol would substantially strengthen the generality claim.
Reviewer_9VUSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work raises an interesting question about the trade-off between privacy-preserving face recognition (PPFR) and face de-identification (De-ID). It proposes a teacher–student framework that trains a student model to produce embeddings that facilitate more effective face reconstruction attacks against both PPFR and De-ID methods.

Critical strengths

Strengths

  • The joint consideration of PPFR and De-ID offers an interesting perspective.
  • The results demonstrate improved attack success.
Critical weaknesses

Weaknesses

  • The theoretical contribution is limited.
  • The trade-off between PPFR and De-ID is insufficiently explored.
  • The rationale underlying the proposed design is insufficiently established.
  • Several aspects of the evaluation setup and consistency of the results require clarification.

Detailed Comments

  1. The paper opens with an interesting discussion of the trade-off between PPFR and De-ID. However, it ultimately presents an attack without adequately addressing this trade-off. In addition, the paper should clarify how the attack findings advance our understanding of the relationship between these two privacy objectives.

  2. The claimed shared structural weakness is not sufficiently characterised. The contribution would be stronger if the paper characterised this weakness at a finer level of detail and supported the analysis with experimental evidence. It should also explain how this weakness motivates the teacher–student framework. For instance, in Section 5.1, the superior performance of the U-Net attack on MinuFace and PartialFace is attributed to “architectural and optimization choices,” without explaining how the framework’s design accounts for this improvement.

  3. Why is a teacher–student framework suitable for this problem? What role does the stop-gradient operator sgsg play in the loss functions, and how important is it to the method’s effectiveness? Why is α\alpha set to 0.8? Please also define p(ϵ)p(\epsilon) on page 5. These choices require clearer explanations and, where appropriate, ablation studies.

4.Page 4 states that a model can learn a mapping from zpz_p back to zmz_m and use the resulting embeddings for linkage. Is this claim supported by existing work? Although the successful attack provides evidence of identity leakage, it does not establish that the proposed method recovers zmz_m. Please directly evaluate the correspondence between the recovered embeddings and zmz_m, including whether fine-grained embedding information is preserved.

  1. The paper uses Face++, the Amazon API, and AdaFace similarity for evaluation. Please explain what each metric measures and why it provides a suitable proxy for human judgements of identity. Why were these measures chosen over other established face similarity metrics? Without human validation or other supporting evidence, it remains unclear how reliably these metrics capture recognisability to human observers.

  2. The statement that impersonation methods cause a face recognition model to identify a protected image as another person while still recognising the actual identity appears contradictory and requires clarification. Does this mean that the actual identity remains among the lower-ranked candidates? If so, this does not necessarily imply little privacy protection. In applications that return only the top-nn matches, excluding the actual identity from those matches may still provide meaningful protection. This reasoning therefore does not justify excluding these methods from the evaluation.

  3. Please justify the choice of MS1M as the source of real images and discuss whether the findings generalise to other real-image datasets. To evaluate transferability against PPFR methods, the paper uses the synthetic TPDNE dataset. Why was another real-image dataset not used? Please discuss whether characteristics of the synthetic data could affect attack performance and the interpretation of the transferability results.

  4. Potential overlap with model training data. How do the authors ensure that the frozen FracFace model was not trained on images or identities included in the attack’s training and test datasets? Please clarify how potential overlap is checked and how it could affect the evaluation results.

  5. Consistency of the reported results. Table 1 reports a value of 0.749 for DecoyFace, whereas Table 4 reports 0.419. Please explain this discrepancy and clarify whether it reflects differences in evaluation settings or instability in the proposed method.

Questions for authors

What is the fundamental methodological novelty that drives the observed improvements, and what evidence distinguishes its contribution from the effects of architecture, optimisation, and dataset choices?

Reviewer_gURiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits identity leakage in privacy-preserving face recognition (PPFR) and perception-preserving face de-identification (De-ID) systems. It proposes FaceLinkGen, a unified knowledge-distillation-based attack that maps protected inputs into a conventional face-identity embedding space using a frozen ArcFace teacher. For PPFR, the distilled embeddings condition Arc2Face to synthesize faces matching the original identity. For De-ID, cross-image identity supervision and an original-domain preservation loss enable linkage between protected and unprotected images. The authors rely on public datasets to show the identity leakage.

Critical strengths
  1. The paper identifies a shared identity-leakage risk in keyless PPFR and perception-preserving De-ID systems, despite their different privacy goals.

  2. FaceLinkGen provides a simple, unified distillation-based attack for both face reconstruction and cross-image identity linkage.

  3. The experiments cover seven representative protection methods and include identity-disjoint evaluation and limited-data settings.

Critical weaknesses
  1. In Section 4, the De-ID module introduces cross-image supervision, original-domain preservation, and the loss weight α=0.8, but their individual effects are not clearly ablated.

  2. The novelty and contribution of the teacher-student model is unclear or limited and should be further enhanced.

  3. In the experimental setup described under Targets and Baseline, most evaluations use identity-disjoint splits of MS1M. Cross-dataset tests on independent real-world face data are missing for De-ID.

  4. In the Metrics section, Face++ and Amazon verification are used as proxies for human recognizability, as reported in Table 1, but no human evaluation validates this assumption.

  5. FaceLinkGen is not always the strongest tested attack. Table 1 shows better results for the tuned U-Net on MinusFace and PartialFace, while Table 2 shows higher Recall@1 for U-Net on TIP-IM.

  6. As described in the Targets and Baseline section, FaceLinkGen requires separate training for different protection methods, and its generalization to unseen protection methods is also worth further investigation.

Questions for authors

See weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.