acceptodds
Under review as a conference paper at ICLR 2027

When Disentanglement Is Not Enough: Identifiability and Semantic Privacy in Learned Communications

Abstract

Advances in representation learning allow for new and previously unfeasible approaches for efficient and secure data transmission over noisy channels. A wireless communication system based on separate source coding, i.e., compression, encryption, and channel coding, i.e., error correcting coding, is proven to be asymptotically optimal and is currently widely used in modern secure wireless networks. Joint source-channel coding is an alternative with a range of advantages over the separation based scheme, including a better communication performance in short data blocklengths and graceful performance degradation with worsening channel quality. Although the integration of encryption, including feature-selective encryption, into the joint source-channel coding is addressed in the literature, it is not studied in the context of capabilities of the existing representation learning models. In this work, we first prove that zero information leakage about the private semantic concept from public latent components is generally not achievable in unsupervised representation learning. We then show that weakly supervised disentangled representation learning models based on variational autoencoders can overcome this limitation and can be used as a base model for semantic factor disentanglement in a joint source-channel-encryption coding.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
disentangled representation learning, representation learning, joint source-channel coding

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

This paper studies semantic privacy in learned joint source-channel-encryption coding (JSCEC), and focuses on whether disentangled representations can prevent private semantic information from leaking through public latent components. The paper shows that unsupervised disentanglement cannot generally guarantee zero leakage due to non-identifiability, while weak supervision with paired observations can provide stronger guarantees under suitable assumptions. For dependent public and private factors, the paper characterizes a lower bound on unavoidable leakage. Building on these observations, this paper proposes a feature-selective JSCEC framework using disentangled representations and keyed Haar random rotations for encryption. Experiments on four disentanglement datasets investigate the relationship between representation quality, reconstruction performance, and semantic privacy.

Critical strengths
  1. The paper addresses an important connection between representation learning and semantic privacy in communication systems. In particular, it highlights that disentanglement alone does not necessarily prevent private semantic information from appearing in public latent components. This is meaningful for feature-selective encryption.

  2. The theoretical analysis provides a useful characterization of semantic privacy under different assumptions. In particular, Proposition 1 establishes a lower bound on unavoidable leakage when public and private factors are dependent, and characterizes conditions under which this bound can be achieved. This provides a principled perspective on the fundamental limits in semantic privacy.

  3. The experiments provide some evidence connecting representation quality to semantic privacy. In particular, the substantial leakage observed on Cars3D, where disentanglement quality is poor, is consistent with the paper’s theoretical motivation.

Critical weaknesses
  1. Limited theoretical novelty. The main theoretical observations appear incremental relative to existing results on unsupervised non-identifiability and weakly supervised disentanglement, both of which are acknowledged and cited by the authors. While their application to semantic privacy is meaningful, the paper does not sufficiently clarify what new theoretical insights are established beyond the existing literature.

  2. Limited discussion of the theoretical assumptions in practical settings. The theoretical guarantees rely on idealized assumptions, including infinite data, exact matching of the data distribution, and specific paired-data conditions. However, the proposed system is implemented using representations learned from finite datasets. The paper provides limited discussion of how these assumptions relate to the practical setting, or how violations of these assumptions may affect semantic privacy.

  3. Insufficient comparison with existing approaches. The paper proposes a feature-selective JSCEC framework but does not provide direct comparisons with existing JSCC/JSCEC methods. The authors also acknowledge this limitation in the conclusion. Without such comparisons, it is difficult to assess the practical advantages and trade-offs of the proposed framework in terms of semantic privacy, reconstruction quality, and communication efficiency.

Questions for authors
  1. How do Theorems 1 and 2 extend the existing results on unsupervised non-identifiability and weakly supervised disentanglement already cited in the paper?

  2. Could the authors further discuss how the idealized assumptions relate to finite-sample training, or provide empirical evidence on how relaxing these assumptions affects semantic privacy?

  3. Could the authors include comparisons with existing JSCC/JSCEC approaches and extend the evaluation to additional representation models or settings?

1 comment
Reviewer_hQgnOpenReview ↗

Official Comment by Reviewer hQgn

Thank you for the clarification and the additional results. Proposition 3 helps connect the theory to the experimental setups, but it is still under the conditions of Proposition 2. My main concern remains how the guarantees and theoretical results relate to practical training when these assumptions are only approximately satisfied or even not. I therefore maintain my original score.

Reviewer_jj97OpenReview ↗
Rating
2(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies semantic privacy in joint source-channel-encryption coding (JSCEC), where private semantic information may leak through public latent components. The authors show that unsupervised disentanglement cannot generally guarantee zero information leakage, while weak supervision can achieve minimum leakage under certain assumptions. They propose a feature-selective JSCEC method using Haar random rotation to encrypt private channel symbols, with information leakage bounded by that of the base representation. The method also considers encryption based on compact Lie groups. Experiments using Ada-GVAE and Lie-VAE on four datasets evaluate reconstruction quality, disentanglement, and the eavesdropper's ability to recover private factors.

Critical strengths
  1. The paper addresses a meaningful problem in feature-selective JSCEC, where private semantic information may leak through public latent components even when private channel symbols are encrypted. The theoretical analysis clarifies the limitations of unsupervised disentanglement and establishes conditions under which weak supervision can reduce information leakage. The derived leakage bound provides a useful connection between representation learning and semantic privacy.

  2. The experiments provide evidence for the feasibility of the proposed JSCEC method. With Ada-GVAE, the eavesdropper's ability to recover private factors remains close to random guessing on three datasets, even when the transmitted signal is observed without channel noise, while the legitimate receiver retains useful information about these factors. These results support the practical relevance of the theoretical analysis, although substantial leakage remains on Cars3D.

Critical weaknesses
  1. The theoretical novelty and its connection to the implemented method are not sufficiently established. The main results on the limitations of unsupervised disentanglement and the benefits of weak supervision build closely on the existing work of Locatello et al. [1] [2]. Although the paper connects these results to information leakage in JSCEC, the additional theoretical contribution remains limited. Moreover, Theorem 2 assumes paired samples with shared public factors and exact distribution matching, while the experiments use Ada-GVAE with a different pair construction and additional labeled samples to identify private components. The connection between the theoretical guarantees and the implemented method is therefore not sufficiently demonstrated.

  2. The experimental evaluation does not sufficiently establish the effectiveness of the proposed method over existing approaches. The paper lacks direct comparisons with closely related feature-selective JSCEC methods, particularly Sun et al. [3], which already addresses private information leakage from public subcodewords. More importantly, the experiments do not isolate the contribution of weak supervision under the same architecture and encryption settings, nor do they directly evaluate information leakage from the public latent components, which is central to the theoretical analysis. Although the reported results demonstrate feasibility on several datasets, they do not clearly establish the benefits of the proposed approach over existing methods or fully validate its main theoretical claims.

References

[1] F. Locatello et al., "Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations," in Proceedings of the 36th International Conference on Machine Learning (ICML), 2019.

[2] F. Locatello et al., "Weakly-Supervised Disentanglement Without Compromises," in Proceedings of the 37th International Conference on Machine Learning (ICML), 2020.

[3] L. Sun, Y. Yang, M. Chen, and C. Guo, "Disentangled Information Bottleneck Guided Privacy-Protective Joint Source and Channel Coding for Image Transmission," IEEE Transactions on Communications, vol. 72, no. 11, pp. 6992–7005, 2024.

Questions for authors
  1. We would like the authors to clarify the additional theoretical contributions of Theorems 1 and 2 beyond the existing results of Locatello et al. [1] [2]. In particular, we would like the authors to explain how the theoretical guarantee in Theorem 2 applies to the implemented Ada-GVAE, which uses a different pair construction and additional labeled samples to identify private latent components.

  2. We would like the authors to provide controlled experiments that isolate the contribution of weak supervision to reducing information leakage from public latent components under the same architecture and encryption settings. We would also like to see direct comparisons with existing feature-selective JSCEC methods, particularly Sun et al. [3], under comparable channel conditions and privacy metrics. These results would help establish whether the proposed method provides meaningful benefits over existing approaches.

Reviewer_AtihOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors prove that unsupervised disentangled representation learning cannot guarantee zero information leakage of private semantic factors into public latent components, even when latent variables are statistically independent. Then the paper provides theoretical guarantees for weakly‑supervised disentanglement using paired samples sharing public factors, showing leakage can be driven down to the inherent mutual information between public and private ground‑truth semantics.

Critical strengths

The paper provides new theoretical insight for semantic privacy. It demonstrates that statistically disentangled independent latents do not guarantee semantic privacy. Even perfectly factorised latents can arbitrarily leak private semantic concepts into public dimensions, which cannot be identified purely from marginal data distribution. This is an important conceptual warning for semantic‑communication community, which often conflates latent statistical independence with semantic separation.

Critical weaknesses

The paper is limited to a restricted threat‑model scope. The eavesdropper is assumed to only probe from observed channel symbols, and knows the decoder weights but not the encryption key. It does not consider adaptive attacks, chosen‑plaintext attacks, or adversarial probing against the latent representation. Moreover, encryption relies on key‑derived pseudorandom generation for Haar matrices.

Questions for authors

All experiments run on synthetic / simulated disentanglement datasets. What is the generalization performance if the proposed JSCEC models are applied to real‑world natural images? Real data exhibits complex factor correlations where public‑private factor dependencies are ubiquitous.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.