Research map / Federated learning and privacy / Communication-efficient FL
Decentralized Federated Learning with Model Caching on Mobile Agents
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 Cached-DFL, a decentralized federated learning framework where mobile vehicles cache and relay models of encountered agents (DTN-style), with a convergence analysis accounting for cache staleness and LRU/group-based cache update rules.
Problem
Decentralized federated learning among mobile agents (e.g., vehicles) suffers slow convergence and low accuracy because sporadic device-to-device contacts limit model spreading, especially under non-i.i.d. data.
Approach
Each mobile agent caches models of agents encountered recently and exchanges its own plus cached models upon meeting (delay-tolerant networking style), then aggregates over all cached models; the paper analyzes convergence with explicit staleness and designs LRU and group-based cache update rules.
Key results
- Cached-DFL with LRU caching converges faster and outperforms DeFedAvg (DFL without caching) on MNIST, FashionMNIST and CIFAR-10 under non-i.i.d., i.i.d. and Dirichlet data, approaching centralized FL especially in non-i.i.d. settings.
- Larger cache sizes and moderate staleness tolerance tau_max improve non-i.i.d. convergence, but larger tau_max lowers final accuracy in both non-i.i.d. and i.i.d. settings.
- Group-based LRU cache update outperforms DFL under all data-overlap settings on FashionMNIST, while vanilla LRU does not outperform DFL at convergence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LRU model caching and aggregation to DFL | DFL · MNIST, FashionMNIST, CIFAR-10; non-i.i.d., i.i.d., and Dirichlet (π=0.5) | test accuracy | Helped |
| Increase cache size (ablation) | LRU Cached-DFL · MNIST and FashionMNIST; especially non-i.i.d. | test accuracy | Helped |
| Increase staleness tolerance τ_max (ablation) | LRU Cached-DFL · MNIST, non-i.i.d. | convergence speed | Helped |
| Increase staleness tolerance τ_max (ablation) | LRU Cached-DFL · MNIST, i.i.d. | test accuracy | No clear effect |
| Increase staleness tolerance τ_max (ablation) | LRU Cached-DFL · MNIST, non-i.i.d. and i.i.d. | final test accuracy | Hurt |
| Increase vehicle mobility speed while reducing local updates to keep wall-clock time comparable (ablation) | LRU Cached-DFL · Non-i.i.d. vehicle network | convergence speed | Helped |
| Replace vanilla LRU with group-based LRU caching (ablation) | Cached-DFL · FashionMNIST; grouped mobility and 0-, 1-, 2-, and 3-overlap data distributions | final test accuracy | Helped |
| Use vanilla LRU caching instead of group-based LRU caching (ablation) | Cached-DFL · FashionMNIST; grouped mobility and data distributions | final test accuracy | Hurt |
Surprising findings
- Vanilla LRU caching converges faster initially but ends up not outperforming DFL at the final stage in the grouped-mobility setting.
- Larger staleness tolerance tau_max reduces final converged accuracy even though it speeds early convergence in non-i.i.d. settings.
Limitations
- Theoretical analysis relies on strong assumptions such as convexity and bounded gradients, and the convergence bound is stated loosely with O-notation.
- Only simple LRU and group-based cache update and aggregation algorithms are studied.
- Evaluation is limited to simulated vehicular mobility (Manhattan model) with small CNN/ResNet-18 models on standard image benchmarks.
Experimental setup
Datasets: MNIST FashionMNIST CIFAR-10
Benchmarks: MNIST FashionMNIST CIFAR-10
Metrics: average test accuracy across vehicles average number and age of cached models
Baselines: DeFedAvg (DFL without caching) Centralized FL (FedAvg)
Builds on: DeFedAvg FedAvg Delay-Tolerant Networking
Closest papers by meaning
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.89)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- Overcoming Data and Model heterogeneities in Decentralized Federated Learning via Synthetic Anchors (ICML 2024 · similarity 0.86)
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.85)
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.85)
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.