Multimodal Decentralized Federated Learning

• Multimodal DFL is a decentralized approach for collaborative model training on heterogeneous data across peer-to-peer networks without a central server.
• It employs methods like PID-guided feature fission and sheaf-theoretic regularization to resolve gradient conflicts and align modality-specific features.
• Empirical evaluations demonstrate that these frameworks improve accuracy, communication efficiency, and robustness under non-IID conditions and modality dropout.

Multimodal decentralized federated learning (DFL) is a paradigm for collaborative model training on heterogeneous multimodal datasets distributed over a peer-to-peer (P2P) network, without central coordination. Each agent or client may possess a different subset of data modalities and model architectures; agents interact via direct peer exchanges to aggregate representational or parameter information, aiming to solve a global multimodal learning task. The inherent challenges arise from modality heterogeneity, architectural diversity, privacy requirements, and absence of a central server, which critically compromise naïve extensions of traditional federated learning (FL) or multimodal representation learning.

1. Foundational Principles and Challenges

The canonical multimodal FL pipeline assumes a single monolithic joint embedding for all modalities, which in decentralized settings causes misalignment of gradients between uni- and multimodal clients and suppresses efficient knowledge transfer. In DFL, clients are not expected to share global parameters or raw data and may differ in the modalities (audio, vision, sensors, text, etc.) and model configurations available. The absence of a central aggregator restricts gradient mixing and peer selection to direct P2P updates, reinforcing the problem of inconsistent update dynamics. The central technical dilemma in multimodal DFL is the resolution of gradient conflict: aggregating gradients from clients with differing modality subsets and objectives often leads to subspace interference and suboptimal convergence (Shi et al., 15 Jan 2026).

2. Information Decomposition in Multimodal DFL

A rigorous theoretical foundation is provided by partial information decomposition (PID), which formalizes the breakdown of joint mutual information $I([X,Y]; Z)$ into redundant $R(Z)$, unique $U_X(Z), U_Y(Z)$, and synergistic $S(Z)$ components. For modalities $X$ and $Y$ and label $Z$,

$I(X, Y; Z) = R(Z) + U_X(Z) + U_Y(Z) + S(Z)$

where $R(Z)$ is information about $Z$ shared by both $X$ and $Y$, $U_X(Z)$ is unique to $X$, $U_Y(Z)$ is unique to $Y$, and $S(Z)$ is that obtainable only when $X$ and $Y$ are observed jointly. In practice, exact PID entropy terms are infeasible for high-dimensional latent codes; approximation schemes compute sample-wise minimum/maximum mutual information for redundancy and synergy, and difference-of-MI terms for uniqueness (Shi et al., 15 Jan 2026).

3. Architectures and Algorithmic Frameworks

Recent advances propose diverse algorithmic frameworks for multimodal DFL, each reflecting different treatments of modality heterogeneity and decentralized topology:

a) PID-Guided Partial Alignment and Feature Fission (PARSE)

PARSE (Shi et al., 15 Jan 2026) implements PID by fissioning each modality's latent vector $z^m$ into three disjoint slices: redundant $(z^{m,r})$, unique $(z^{m,u})$, and synergistic $(z^{m,s})$. These slices are mixed among peers restricted by modality availability:

• Redundant slices are globally aligned via decentralized contrastive loss ($\mathcal{L}_{\text{nce}}$) exchanged only among agents sharing the corresponding modality.
• Unique and synergy slices remain local, with synergy slices mixed within subgraphs of agents possessing identical modality subsets.

The key workflow per agent (in summary form):

1. Local SGD step on all slices.
2. Neighbor parameter mixing only for redundant slices among shared modality peers.
3. Multimodal agents mix synergy-head parameters within exclusive subgraphs.

No gradient surgery is necessary; the separation ensures subspace-orthogonality and resolves cross-task parameter conflicts. Full agent objectives involve supervised classification on redundant/unique slices, contrastive alignment for redundant slices, and synergistic fusion/classification for multimodal agents (Shi et al., 15 Jan 2026).

b) Sheaf-Theoretic Collaboration (Sheaf-DMFL, Sheaf-DMFL-Att)

Sheaf-DMFL (Ghalkha et al., 27 Jun 2025) employs cellular sheaf theory to regularize pairwise consistency among client-specific task-head projections. Feature encoders per modality are collaboratively trained among clients holding that modality; concatenated or attention-fused client heads are regularized using a sheaf Laplacian:

$$\min_{\{\theta_i\}} \sum_{i=1}^N f_i(\theta_i) + \frac{\lambda}{2} \theta^\top L_\mathcal{F} \theta$$

where $L_\mathcal{F}$ is the cellular sheaf Laplacian constructed from restriction maps $P_{ij}$ across the communication graph. Sheaf-DMFL-Att enhances fusion via a softmax-attention over modality embeddings, with mathematically established $O(1/R)$ convergence rate (Ghalkha et al., 27 Jun 2025).

c) Distillation-based Embedding Knowledge Transfer (FedMEKT)

FedMEKT (Le et al., 2023) uses a semi-supervised, server-coordinated protocol: each client trains local multimodal autoencoders and transmits only embeddings (not weights) on a small public proxy set; the server distills these into global encoders via knowledge transfer loss. Subsequent classifier heads are trained atop the distilled encoders, fusing modal representations by concatenation, under both supervised and reconstruction losses (Le et al., 2023). The communication cost is reduced, as only low-dimensional embeddings are exchanged, not full weight matrices.

d) Personalized Local Model Aggregation and Alignment (FedEPA)

FedEPA (Zhang et al., 16 Apr 2025) integrates learnable, element-wise personalized model aggregation weights, adapting global/local parameters for each client. Modality alignment is performed by decomposing features into "aligned" (cross-modal shared) and "context" (modality-private) components. Aligned features are encouraged to be similar across modalities via InfoNCE-style contrastive loss; independence and diversity among aligned/context features are driven by HSIC and Jensen-Shannon divergence penalties. Feature fusion applies modality-wise self-attention over aligned features, and classification is performed on the integrated embedding. The algorithm empirically mitigates client drift and label scarcity, scaling robustly to non-IID partitioning and complex modality splits (Zhang et al., 16 Apr 2025).

4. Optimization Objectives and Data Flow

Typically, decentralized multimodal FL frameworks optimize composite objectives that combine supervised cross-entropy, contrastive alignment, independence penalties, and knowledge distillation losses. For example, the full FedEPA objective is:

$$\min_{\Theta, \Psi} \sum_{i=1}^N p_i \Big[ \mathcal{L}_{\mathrm{sup}}^i(\Theta,\Psi) + \alpha \mathcal{L}_{\mathrm{con}}^i(\Theta) + \beta \mathcal{L}_{\mathrm{HSIC}}^i(\Theta) + \gamma \mathcal{L}_{\mathrm{JSD}}^i(\Theta) \Big]$$

where $\Theta, \Psi$ are encoder and classifier parameters; $\mathcal{L}_{\mathrm{sup}}$ is the supervised loss; $$\mathcal{L}_{\mathrm{con}}, \mathcal{L}_{\mathrm{HSIC}}, \mathcal{L}_{\mathrm{JSD}}$$ are unsupervised alignment and regularization losses.

Communication protocols vary: PARSE and Sheaf-DMFL rely on direct peer updates governed by graph topology and client modality sets; FedMEKT operates under a lightweight server coordinating embedding exchanges, reducing uplink volume to a fraction of naive FL (Le et al., 2023).

5. Empirical Evaluation and Benchmarks

Experimental validation spans image, audio-visual, sensor, and text benchmarks:

Framework	Datasets Used	Key Metrics	Empirical Gains
PARSE (Shi et al., 15 Jan 2026)	KU-HAR, ModelNet-40, AVE, IEMOCAP	Test accuracy per agent	+1–2 pp multimodal accuracy; +0.5–1 pp unimodal
FedMEKT (Le et al., 2023)	UCI HAR, PAMAP2, MHEALTH	Acc., F1-score	+3–7% accuracy over FedAvg; 4–10× less communication
Sheaf-DMFL (Ghalkha et al., 27 Jun 2025)	mmWave blockage, beamforming	Test acc./convergence	Fastest convergence, robust to "partial view"
FedEPA (Zhang et al., 16 Apr 2025)	MGCD, Weibo, UTD-MHAD	OA, BA, F1	>20 pp improvement (MGCD); >38 pp (UTD-MHAD)

Across studies, multimodal DFL consistently outperforms unimodal and hybrid FL baselines, maintains robustness under non-IID and heterogeneous splits, and demonstrates scalability to large peer networks and variable modality availability. Attention and contrastive alignment-based methods (FedEPA, Sheaf-DMFL-Att) yield the strongest results in regime of limited labels and noisy modalities.

6. Modality Alignment, Personalization, and Gradient Conflict Resolution

A central technical insight is that representation fission (as in PARSE) or decompositional alignment (as in FedEPA) avoids destructive gradient mixing by routing unique, redundant, and synergistic information through separate peer update channels. Empirical ablations confirm:

• Balanced latent splits (equal dimensions to redundant, unique, synergistic) maximize joint accuracy (Shi et al., 15 Jan 2026).
• Fusion strategies that use mean or attention-based combination of synergy slices deliver near-optimal performance with minimal communication/parameter cost (Shi et al., 15 Jan 2026, Ghalkha et al., 27 Jun 2025).
• Personalized aggregation weights suppress non-IID drift and improve stability (Zhang et al., 16 Apr 2025).

No explicit gradient surgery is required in these frameworks, as orthogonal subspace assignment via feature slicing or head regularization produces semantically consistent alignment (Shi et al., 15 Jan 2026).

7. Applications and Future Directions

Multimodal DFL frameworks are validated in wireless link prediction, multimedia classification, emotion recognition, and human activity monitoring scenarios. Sheaf-DMFL and PARSE set the stage for further investigation of topological regularization, task-specific peer graphs, and adaptive modality fusion mechanisms. Current evidence demonstrates convergence rates competitive with centralized and server-based approaches and resilience to modality dropout, adversarial splits, and distribution shifts (Shi et al., 15 Jan 2026, Ghalkha et al., 27 Jun 2025).

A plausible implication is that continued advances will further integrate communication-efficient protocols with principled information decomposition, accommodating dynamic agent mixes and more complex real-world multimodal federated deployments. Robustness to non-IID client data, modality loss, and private information protection remain central themes for ongoing and future research in decentralized collaborative learning.