Private-Shared Latent Disentanglement

• Private-shared latent disentanglement is the process of partitioning latent spaces into universal (shared) and idiosyncratic (private) factors for clearer, more interpretable modeling.
• It supports robust applications in federated learning, multimodal analysis, and neuroscience by separating common global patterns from modality-specific or client-specific variations.
• Recent advances include dual-encoder VAEs, adversarial regularization, and gradient gating techniques that optimize reconstruction fidelity and disentanglement quality.

Private-shared latent disentanglement refers to machine learning frameworks that factorize a latent variable space into subspaces corresponding to “shared” (universal, cross-group, or cross-modal) and “private” (group-specific, modality-specific, or client-specific) factors of variation. This concept is foundational across multimodal representation learning, federated learning, neuroscience, and fairness, enabling interpretable and robust learning by explicitly distinguishing universal structure from idiosyncratic or sensitive attributes.

1. Formal Definition and Taxonomy

Let $\mathbf{x}^{(m)}$ denote observed data from modality, client, or group $m \in \{1,\dots,M\}$, and $z_s$ the shared latent, with $z_p^{(m)}$ the private latent of $m$. Generative models typically assume the following structured factorization: $$p\bigl(\{\mathbf{x}^{(m)}\}, z_s, \{z_p^{(m)}\}\bigr) = p(z_s) \prod_{m=1}^M p(z_p^{(m)})\, p_\theta(\mathbf{x}^{(m)}\,|\,z_s, z_p^{(m)})$$ Each $\mathbf{x}^{(m)}$ is reconstructed from both the global $z_s$ and its own $z_p^{(m)}$. The independence prior $p(z_s) \prod_m p(z_p^{(m)})$ enforces non-redundant allocation of signal between subspaces, foundational to effective disentanglement (Märtens et al., 2024, Lee et al., 2020).

Variants are distinguished by whether $m \in \{1,\dots,M\}$0 is continuous or discrete (Lee et al., 2020, Choi et al., 2020), whether $m \in \{1,\dots,M\}$1 equals the number of clients, regions, or modalities (Yan et al., 2023, Soroushmojdehi et al., 28 Oct 2025), and by specific statistical priors (e.g., Gaussian, mixture-of-Gaussians, Gumbel-Softmax). Discrete private latents often model class-specific variation, while continuous privatized subspaces represent style, bias, or temporal context.

2. Model Architectures and Training Objectives

Most modern approaches utilize autoencoder architectures with partitioned latent spaces, employing dedicated encoder heads for $m \in \{1,\dots,M\}$2 and $m \in \{1,\dots,M\}$3 (or $m \in \{1,\dots,M\}$4 for class-conditional private factors). Canonical instantiations are:

• Dual-encoder VAEs for federated clients (Yan et al., 2023), where $m \in \{1,\dots,M\}$5 (shared/global) and $m \in \{1,\dots,M\}$6 (private/personalized) are encoded by two separate encoders, and the local decoder consumes $m \in \{1,\dots,M\}$7.
• Crossed autoencoders for multi-view data, where four encoders (private and shared per modality) are paired with cross-modality decoders, enforcing that each modality’s reconstruction path leverages the other view’s shared codes to block information leakage (Koukuntla et al., 2024).
• Multi-encoder multi-decoder autoencoders with alignment modules and orthogonality losses for multi-region brain signals (Soroushmojdehi et al., 28 Oct 2025).
• Post-hoc factorization VAEs for pretrained embeddings, such as ID/attribute disentanglement in face recognition (Öztürk et al., 13 Apr 2026).

The core objective universally combines:

• Reconstruction loss: Ensuring each $m \in \{1,\dots,M\}$8 is reconstructible from its private and shared codes.
• KL or independence losses: Enforcing that latents conform to their priors and penalizing correlations (e.g., via total correlation or variance-based disentanglement isolates).
• Adversarial or mutual information losses: Minimizing predictability of shared information from private codes, or more generally, controlling $m \in \{1,\dots,M\}$9 (Koukuntla et al., 2024, Öztürk et al., 13 Apr 2026).

Notably, the implementation of the product-of-experts for multimodal shared posteriors enables robust inference when modalities are missing (Lee et al., 2020, Märtens et al., 2024).

3. Disentanglement Mechanisms and Theoretical Properties

The disentanglement of private and shared latents relies on explicit architectural and loss-function design:

• Statistical independence via prior factorization: Imposing $z_s$0 precludes shared-private redundancy.
• Total correlation (TC) penalization: Up-weighting the TC term in the KL-divergence objective encourages independent subfactors within and across private/shared blocks (Lee et al., 2020).
• Predictability minimization (adversarial disentanglement): Minimizing the variance of auxiliary predictors tasked with inferring $z_s$1 from $z_s$2 provides first-order guarantees of mutual information minimization (Koukuntla et al., 2024).
• Capacity/blocked gradients: Gradient gating ensures that same-view reconstructions update only private encoders, while cross-view reconstructions update only shared encoders—robustly separating signals even as private-feature dimensionality becomes large (Märtens et al., 2024).
• Alignment and orthogonality regularization: Penalties on cross-covariance between shared and private latents further suppresses subspace interference, as implemented in SPIRE (Soroushmojdehi et al., 28 Oct 2025).

Absence of strong identifiability guarantees remains a limitation; disentanglement quality is assessed empirically via cross-modal prediction $z_s$3, AUC of latent-space classifiers, and linear separability (Märtens et al., 2024, Öztürk et al., 13 Apr 2026).

4. Empirical Results and Comparative Performance

Empirical studies consistently demonstrate the importance of explicit private-shared factorization:

• Federated Learning: In non-i.i.d. federated settings, dual-encoder VAEs (FedDVA) achieve superior personalization and reduced client variance relative to baseline FL and personalization techniques, yielding absolute accuracy gains of 1.3–2.0% and accelerated convergence (Yan et al., 2023).
• Multimodal and Multi-view Data: On tasks with dominant modality-specific variation, methods incorporating cross-modal reconstruction and gradient gating (MMVAE++, SPLICE) outperform traditional MVAEs, robustly recovering shared factors and achieving high cross-modal $z_s$4 even when private features vastly outnumber shared ones (Märtens et al., 2024, Koukuntla et al., 2024). Geometry-preserving objectives further enable accurate recovery of nonlinear submanifold structures.
• Semi-supervised and Cross-domain Inference: DMVAE hybridizes discrete and continuous private/shared latents, enabling strong semi-supervised classification performance with minimal paired data, and producing qualitatively distinct interpolations under latent space traversal (Lee et al., 2020). Discond-VAE achieves state-of-the-art conditional disentanglement in synthetic and real datasets, cleanly separating class-conditional continuous variation (Choi et al., 2020).
• Sensitive Attribute Suppression: VLEED demonstrates a controllable privacy–utility trade-off frontier in post-hoc face embedding debiasing, outperforming linear nullspace and dimension-elimination baselines, with effective reduction of nonlinear leakage and group-disparate error (Öztürk et al., 13 Apr 2026).

5. Methodological Advances and Variations

Key innovations for robust private-shared disentanglement include:

• Geometry-preserving mappings: SPLICE estimates geodesic distances on projected manifolds and constrains latent-space Euclidean distances to preserve data geometry, a feature critical for interpretable nonlinear factorization (Koukuntla et al., 2024).
• Mutual-information–maximization adversaries: VLEED directly optimizes the conditional entropy of the sensitive variable in the residual (private) latent, outperforming traditional moment-matching and nullspace projection approaches (Öztürk et al., 13 Apr 2026).
• Federated personalized representations: FedDVA's integration with FedAvg enables modular plug-in to FL protocols, synchronizing only encoders while keeping decoders local, balancing global knowledge sharing with individual customization and yielding interpretable factors inspectable via synthetic interventions (Yan et al., 2023).
• MMVAE++ gradient flow modifications: Addressing the issue of shared-latent “collapse” under overwhelming private signal by strictly separating update paths in the ELBO objective (Märtens et al., 2024).

The following table summarizes several representative approaches:

Method	Domain/Setting	Disentanglement Mechanism
FedDVA	Federated learning	Dual encoder (global/private), FL
SPLICE	Multi-view/neuro	Crossed AE, adversarial, geometry
DMVAE	Multimodal, semi-sup	PoE posterior, $z_s$5-TCVAE loss
MMVAE++	Multi-omics	MoE, explicit gradient gating
VLEED	Face recognition	Post-hoc VAE, entropy minimization
SPIRE	Multi-region neural	AE with alignment/ortho losses

6. Application Domains and Interpretability

Private-shared latent disentanglement is broadly applicable:

• Federated learning: Enhanced personalization and explainability, enabling global knowledge transfer and client privacy (Yan et al., 2023).
• Neuroscience: Functional network modeling of brain regions, attributing variance to shared dynamics or region-specific signatures (Soroushmojdehi et al., 28 Oct 2025, Koukuntla et al., 2024).
• Multi-omics and biomedicine: Identification of disease-relevant molecular signatures that generalize across modalities, robust to high-dimensional private features (Märtens et al., 2024).
• Fairness, privacy, and debiasing: Post-hoc separation of sensitive attributes (gender, ethnicity) from identity, improving group fairness in biometric systems (Öztürk et al., 13 Apr 2026).
• Semi-supervised learning and conditional generation: Superior label efficiency via explicit cross-view transfer, style transfer by latent traversal, and interpretable latent variable manipulation (Lee et al., 2020, Choi et al., 2020).

Interpretability is enhanced via latent traversals, t-SNE clustering of subspaces, and linear probe diagnostics, clarifying the attribution of performance gains to specific latent factors.

7. Best Practices, Limitations, and Open Problems

Reliable private-shared disentanglement requires architectural partitioning, robust cross-modal/cross-client objectives, and independence-enforcing penalties. Best practices include empirical monitoring of cross-modal $z_s$6, calibration of latent dimensionalities, and—where necessary—the inclusion of minimal label supervision to resolve ambiguities in shared structures.

Limitations include the absence of theoretical identifiability under general nonlinear settings, sensitivity to hyperparameter and architecture selection, and open challenges in extending to $z_s$7 settings, hierarchical or nested latent structures, and complex covariate correction (Märtens et al., 2024, Choi et al., 2020).

Future research directions encompass principled uncertainty quantification, adaptive discovery of latent dimensionality, information-theoretic guarantees of disentanglement, joint multi-attribute suppression, and effective generalization to arbitrary multimodal and distributed settings.