acceptodds
Under review as a conference paper at ICLR 2027

Learning Privacy-Preserving Representations for Audio-Visual Deepfake Detection

Abstract

We introduce Privacy-Preserving Audio-Visual Deepfake Detection (PPDD), which aims to suppress private information while preserving robust and generalizable deepfake detection. Existing privacy-preserving approaches primarily target unimodal tasks and are inadequate for audio-visual deepfake detection, where subtle spatial and temporal manipulation cues may coexist with privacy information, requiring fine-grained privacy suppression. To address this, we propose an end-to-end adversarial framework that learns privacy-preserving audio and visual representations using modality-specific anonymizers. A Feature Suppression Block (FSB) performs fine-grained feature-level suppression, guided by unimodal and cross-modal privacy supervision. Experiments on FakeAVCeleb demonstrate substantial privacy suppression with negligible loss in detection utility. The learned representations further preserve detection utility across unseen manipulations and datasets, transfer across detectors and modalities, suppress unseen privacy attributes, and remain resistant to inversion attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Privacy Preservation, Deepfake Detection, Audio-Visual

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

This paper studies privacy-preserving audio-visual deepfake detection. The authors propose separate audio and video anonymizers, which are trained against a proxy privacy attacker while preserving deepfake detection performance. The key component is a feature suppression block (FSB) that suppresses features within each token instead of dropping entire tokens. Experiments show that the proposed method reduces attribute leakage with little loss in detection performance.

Critical strengths
  1. The paper is well motivated: Deepfake detection often requires access to facial videos and speech, which may also expose sensitive information about the user. I find it practically useful to study how this information can be protected while keeping the detector effective. This is especially relevant when users submit their recordings to a third-party detection service.
  2. The proposed method achieves an encouraging privacy-utility balance, in terms of privacy cMAP and detection AUC in Tab. 1.
Critical weaknesses
  1. The scope of privacy protection could be clearer: The paper measures demographic (i.e., gender, ethnicity, age) leakage, but does not directly test identity leakage that whether a person can still be recognized from the outputs. I understand that the authors cannot cover every privacy risk; However, identity is relevant to the threats mentioned in the paper’s introduction (e.g., identity theft).
  2. The method is intuitive, but its novelty and analysis seem relatively limited: Adversarial privacy learning (against a proxy attacker model) is already widely established, and FSB mainly adds feature-wise gating. The ablation supports its benefit, but does not explain what it learns to remove or preserve.
Questions for authors
  1. Could the authors clarify whether the privacy goal includes hiding identity? If so, a face or speaker recognition test would be helpful. If the goal is limited to the evaluated attributes, I would suggest making this clearer in the paper’s introduction.
  2. Could the authors discuss the performance drop on KoDF relative to NFFP? It would also be useful to know whether the privacy benefit remains when the attacker is trained on anonymized KoDF data.
  3. A common concern with adversarial privacy learning is that the protection may depend on the chosen proxy attacker. I appreciate that the authors retrain the attacker after freezing the anonymizers. However, the attacker still uses the same architecture. Could the authors test an alternative privacy attacker to see whether the learned protection transfers?
Reviewer_hkFpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

To suppress private information while preserving robust and generalizable deepfake detection, the paper proposes modality-specific anonymizers with a Feature Suppression Block (FSB) that performs fine-grained feature-level suppression of privacy-related information while preserving deepfake cues. Experiments on FakeAVCeleb report substantially reduced privacy-attribute prediction performance with almost unchanged intra-dataset detection AUC. Additional experiments examine cross-manipulation and cross-dataset generalization, detector transfer, unseen privacy attributes, and inversion attacks.

Critical strengths
  • The paper addresses the privacy–utility problem. Deepfake detection relies on facial and speech information that can also reveal sensitive attributes, making indiscriminate obfuscation potentially harmful to detection. The modality-specific anonymizers and joint privacy supervision provide a reasonable approach to this problem. Table 1 shows a promising empirical trade-off that average privacy cMAP decreases from 93.54 to 41.96, while detection AUC changes from 96.30 to 96.24.
Critical weaknesses
  • It seems that the methodological novelty is incremental. Existing work (as related work section) already explores adversarial feature anonymization and recognizes that privacy and task-relevant information may coexist. The paper mainly adapts these ideas to audio-visual deepfake detection through feature-wise gating, modality-specific anonymizers, and joint privacy supervision. The combination is well motivated, but its conceptual advance over existing privacy-preserving representation learning remains limited.
  • The reduction in deepfake detection is not consistently negligible. While intra-dataset AUC remains almost unchanged, Table 3(a) reports 95.8 AUC from FakeAVCeleb on KoDF introducing PPDD (compared with 100.0 AUC of NFFP and 99.0 AUC of ICSAVDF). The paper should explicitly acknowledge the degradation and qualify its claims of preserving detection utility.
Questions for authors
  • It seems that the reliability of privacy protection is not sufficiently established. The main evaluation uses the same architecture employed during adversarial training. It remains unclear whether sensitive attributes could be recovered using different attacker architectures? In addition, the main results in Table 1 aggregate multiple attributes (gender, ethnicity, age) without per-attribute results or random-choice baselines, making it difficult to assess which attributes remain exposed.
Reviewer_NqGpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Summary:

This paper proposes PPDD for privacy-preserving audio-visual deepfake detection. It uses audio and video anonymizers with a Feature Suppression Block (FSB), trained adversarially to reduce privacy leakage while retaining deepfake detection performance. The experiments show good privacy-utility results and evaluate generalization across several settings.

Critical strengths

Strength:

The paper studies an important problem, as audio-visual deepfake detection often relies on face and voice information that may also reveal private information.

The method achieves a promising privacy-utility trade-off, with strong detection performance and reduced privacy leakage. The evaluation also covers several generalization settings.

Critical weaknesses

weakness:

The privacy evaluation focuses on five attributes: gender, ethnicity, age, valence, and arousal. However, facial and speaker identity are important sources of privacy leakage in audio-visual data, particularly in the context of deepfake detection. The paper does not directly evaluate whether the anonymized representations can still be used for face or speaker recognition. This leaves a key aspect of the claimed privacy protection unverified.

The authors use the same privacy attacker architecture for both training and evaluation. It is unclear whether the method actually suppresses private information or only makes it difficult for this particular model to extract. Privacy attackers with different architectures may still be able to recover the same sensitive attributes.

In the experimental setup, the privacy evaluation uses 64-frame clips, while deepfake detection uses 256-frame clips. This gives the detector more temporal information than the privacy attacker. Since longer sequences may also reveal additional private information, it is unclear whether the reported privacy-utility trade-off would hold when both are evaluated with the same amount of information.

Questions for authors

See weaknesses

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for AC

Dear Reviewers,

We thank all reviewers for their careful reading and constructive feedback. We hope our responses and revisions have addressed the concerns raised.

Dear Area Chair and Meta-Reviewer,

Thank you for handling our submission. Below, we summarize the key reviewer concerns, our responses, and the corresponding manuscript revisions.

Major Concerns

1. Privacy Suppression Against Different Attacker Architectures

Concern: Whether privacy suppression generalizes beyond the attacker architecture used during training.

Our response:

  • We evaluated alternative privacy attackers using CNN-based audio/video encoders and CNN/Transformer multimodal fusion.
  • Privacy leakage remains substantially lower across these architectures, demonstrating generalization beyond the original attacker.
  • Manuscript: Appendix A.4.4.

2. Identity Leakage Through Face and Speaker Recognition

Concern: Whether PPDD suppresses identity information beyond the attributes used during training.

Our response:

  • We evaluated FaceNet and ECAPA-TDNN on 300 VoxCeleb2 identities.
  • Face recognition Top-1 accuracy dropped from 98.56% to 0.33%, and speaker recognition from 98.56% to 0.89%, demonstrating substantial identity suppression.
  • Manuscript: Appendix A.4.5.

3. Methodological and Task-Level Novelty

Concern: Limited methodological novelty given existing adversarial privacy learning and feature suppression techniques.

Our response:

  • While adversarial privacy learning and token-level suppression are established, directly adapting existing privacy-preserving action recognition methods may be insufficient for deepfake detection, because manipulation and privacy cues can coexist within the same tokens. This motivates our fine-grained Feature Suppression Block (FSB).
  • Our framework jointly learns unimodal and multimodal privacy suppression while supporting unimodal detection through independently anonymized audio/video.
  • Although FSB is conceptually simple, it incorporates multiple design choices, including progressive dimensionality reduction, suppression regularization, and λ\lambda scheduling, whose contributions are demonstrated through ablations.
  • We also emphasize the task-level novelty of privacy-preserving audio-visual deepfake detection, which, to the best of our knowledge, has not been explored previously.
  • Manuscript: We have incorporated these points into the revised Introduction.

Additional Concerns

4. Cross-Dataset Privacy Generalization

Concern: Whether privacy suppression transfers to a different dataset.

Our response:

  • We evaluated FakeAVCeleb-trained anonymizers using privacy attackers trained on raw and anonymized CREMA-D data.
  • Average privacy cMAP decreased from 68.16% to 41.84%, demonstrating cross-dataset privacy suppression.
  • Manuscript: Appendix A.4.3, Table 8, alongside KoDF results.

5. Cross-Dataset Detection Performance Drop

Concern: The 4.2 percentage-point AUC reduction is not negligible.

Our response:

  • We acknowledge this limitation and discuss the privacy-utility trade-off. While this remains a limitation, we also evaluate generalization across other dimensions, including unseen manipulations, unimodal detection, and different deepfake detectors.
  • Manuscript: Main paper (lines 454–461).

6. Attribute-Level Privacy Suppression

Concern: Privacy suppression across individual attributes and comparison with random-choice baselines.

Our response:

  • We report attribute-level privacy cMAP for gender, ethnicity, age, valence, and arousal, alongside random-choice baselines.
  • Manuscript: Appendix A.4.2.

Closing Remarks

Across the reviews, several key concerns centered around the generalization of the learned anonymization beyond the original privacy attributes (identity suppression), attacker architecture, and dataset. We have addressed these through additional experiments and incorporated the findings into the revised manuscript.

We have also clarified our methodological and task-level novelty and acknowledged the cross-dataset detection limitation.

We thank the reviewers, Area Chair, and Meta-Reviewer for their time and consideration. We hope this summary is helpful in assessing our revised submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.