Research map / Federated learning and privacy / Personalized and heterogeneous FL
On Disentanglement of Asymmetrical Knowledge Transfer for Modality-Task Agnostic Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formulates modality-task agnostic federated learning (AFL) and proposes DisentAFL, which uses two-stage knowledge disentanglement and gating to split asymmetrical inter-client knowledge sharing into symmetrical sub-schemes.
Problem
Federated learning across clients with different input modalities, downstream tasks, domains and concepts creates asymmetrical pairwise knowledge relationships, making it hard to maximize positive transfer and minimize negative transfer between clients.
Approach
Disentangles the asymmetrical inter-client information-sharing scheme into several independent symmetrical schemes via coarse group-wise (encoder/decoder) and fine-grained knowledge-type disentanglement, realized with a supernetwork of mixture-of-experts layers (MoDE, MoME, MoTE), disentanglement losses, and sparsely-gated client routing.
Key results
- On MERGE-VM, DisentAFL-Avg reaches 96.38% vs 90.11% for the best baseline Cross-PFL.
- On ModelNet-xM, DisentAFL-KD reaches 96.62% vs 94.20% for Cross-PFL.
- Ablation on MERGE-FA: DisentAFL-Avg drops from 93.30% to 85.56% without stage-one auxiliary loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DisentAFL-Avg instead of Cross-PFL | AFL clients · MERGE-AC | accuracy: 81.58 ± 0.53 -> 82.66 ± 0.74 | Helped |
| Use DisentAFL-Avg instead of Cross-PFL | AFL clients · Vehicle-xM | accuracy: 86.82 ± 0.38 -> 88.56 ± 0.22 | Helped |
| Use DisentAFL-KD instead of Cross-PFL | AFL clients · ModelNet-xM | accuracy: 94.20 ± 0.25 -> 96.62 ± 0.16 | Helped |
| Use DisentAFL-Avg instead of Cross-PFL | AFL clients · MERGE-VM | accuracy: 90.11 ± 0.63 -> 96.38 ± 0.41 | Helped |
| Use DisentAFL-Avg instead of Cross-PFL | AFL clients · MERGE-MM | accuracy: 75.37 ± 0.26 -> 75.68 ± 0.74 | No clear effect |
| Add the stage-one auxiliary loss (ablation) | DisentAFL-KD · MERGE-FA | accuracy: 88.44 ± 0.36 -> 91.51 ± 0.36 | Helped |
| Add the stage-two auxiliary loss (ablation) | DisentAFL-KD · MERGE-FA | accuracy: 87.93 ± 0.93 -> 91.51 ± 0.36 | Helped |
| Add the stage-one auxiliary loss (ablation) | DisentAFL-Avg · MERGE-FA | accuracy: 85.56 ± 0.31 -> 93.30 ± 0.25 | Helped |
| Add the stage-two auxiliary loss (ablation) | DisentAFL-Avg · MERGE-FA | accuracy: 86.48 ± 0.41 -> 93.30 ± 0.25 | Helped |
Surprising findings
- Naive cross-group sharing (Cross-FedAvg, Align-FedAvg) often underperforms plain Local training on some simulations, indicating negative transfer.
Limitations
- Evaluated only on simulated AFL setups built from existing multimodal/multitask datasets, with at most 4 modalities and 4 tasks.
- Requires predefined numbers of domains and modality/task indicators for gating.
Experimental setup
Datasets: MERGE dataset simulations ModelNet40 (multi-view) Vehicle driving dataset MNIST/Omniglot-style image classification (Finn et al.) audio-image classification dataset (Liang et al.)
Benchmarks: MERGE-AC Vehicle-xM ModelNet-xM MERGE-VM MERGE-MM MERGE-FA
Metrics: average testing accuracy over clients
Baselines: Local FedAvg Cross-FedAvg Align-FedAvg Cross-PFL FedMSplit
Builds on: FedAvg FedMSplit personalized federated learning federated knowledge distillation
Closest papers by meaning
- Bridging Model Heterogeneity in Federated Learning via Uncertainty-based Asymmetrical Reciprocity Learning (ICML 2024 · similarity 0.86)
- Generalization in Federated Learning: A Conditional Mutual Information Framework (ICML 2025 · similarity 0.85)
- Towards Diverse Device Heterogeneous Federated Learning via Task Arithmetic Knowledge Integration (NeurIPS 2024 · similarity 0.85)
- Complementary Knowledge Distillation for Robust and Privacy-Preserving Model Serving in Vertical Federated Learning (AAAI 2024 · similarity 0.85)
- Data-Free Black-Box Federated Learning via Zeroth-Order Gradient Estimation (AAAI 2025 · similarity 0.84)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.