Research map / Federated learning and privacy / Personalized and heterogeneous FL
DCHM: Dynamic Collaboration of Heterogeneous Models Through Isomerism Learning in a Blockchain-Powered Federated Learning Framework
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.
Proposes DCHM, a blockchain-powered federated framework that coordinates heterogeneous ZNN solver models via hierarchical aggregation (DHA) to track time-varying body-shape changes.
Problem
Centralized RNN/ZNN models and standard federated learning struggle with time-varying problems, model heterogeneity, and reliance on a central server; the paper targets distributed, blockchain-coordinated solving of time-varying linear equations for body-shape analysis.
Approach
Heterogeneous ZNN solver models run on permissionless nodes within groups, with permissioned nodes aggregating results (weighted by inverse error) and secondarily aggregating across groups over a consortium blockchain; only computed results, not parameters, are shared.
Key results
- DCHM average MAE 3.29e-2 and RMSE 3.53e-2, lower than Model2 (4.55e-2), Model3 (26.29e-2) and Model4 (14.93e-2) but slightly higher than Model1 (2.30e-2).
- Average errors over selected rounds: DCHM 1.3835e-2 vs Model1 2.0074e-2, Model2 4.0694e-2, Model3 15.4802e-2, Model4 15.079e-2.
- Per-variable deviations of DCHM on the body-shape equation are 0.0363, 0.0266, 0.0389, 0.0511.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCHM instead of Model1 | Body shape temporal analysis · Body shape temporal analysis | MAE: 2.30 × 10−2 vs 3.29 × 10−2 | Hurt |
| Use DCHM instead of Model1 | Body shape temporal analysis · Body shape temporal analysis | RMSE: 2.31 × 10−2 vs 3.53 × 10−2 | Hurt |
| Use DCHM instead of Model2 | Body shape temporal analysis · Body shape temporal analysis | MAE: 4.55 × 10−2 vs 3.29 × 10−2 | Helped |
| Use DCHM instead of Model2 | Body shape temporal analysis · Body shape temporal analysis | RMSE: 4.58 × 10−2 vs 3.53 × 10−2 | Helped |
| Use DCHM instead of Model3 | Body shape temporal analysis · Body shape temporal analysis | MAE: 26.29 × 10−2 vs 3.29 × 10−2 | Helped |
| Use DCHM instead of Model3 | Body shape temporal analysis · Body shape temporal analysis | RMSE: 27.47 × 10−2 vs 3.53 × 10−2 | Helped |
| Use DCHM instead of Model4 | Body shape temporal analysis · Body shape temporal analysis | MAE: 14.93 × 10−2 vs 3.29 × 10−2 | Helped |
| Use DCHM instead of Model4 | Body shape temporal analysis · Body shape temporal analysis | RMSE: 14.93 × 10−2 vs 3.53 × 10−2 | Helped |
Surprising findings
- The proposed DCHM is not best on MAE/RMSE; Model1 outperforms it, which the authors attribute to the decentralized weight distribution scheme.
Limitations
- Evaluation uses a single self-collected 90-day image sequence and a toy 2x2 linear equation, with no standard benchmark.
- Does not beat Model1 on MAE/RMSE; the authors attribute the gap to decentralized weight distribution.
- Claims about LLM/distributed LLM benefits are not supported by experiments.
Experimental setup
Datasets: self-collected fitness progress photos over 90 days
Benchmarks: self-collected body-shape time-varying linear equation (2x2 matrices, t in [0,30]s)
Metrics: deviation mean absolute error (MAE) root mean square error (RMSE) steady-state error
Baselines: Model1 (Jin et al. 2018) Model2 (Wang et al. 2022) Model3 (Zhang, Ma, and Cai 2008) Model4 (Wang et al. 2021)
Builds on: federated averaging (FedAvg) practical Byzantine fault tolerance (PBFT) zeroing neural network
Closest papers by meaning
- Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization (NeurIPS 2025 · similarity 0.84)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.84)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.83)
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 2025 · similarity 0.83)
- FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning (AAAI 2024 · similarity 0.83)
- FedMBridge: Bridgeable Multimodal Federated Learning (ICML 2024 · similarity 0.83)
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.