FLEX-SFL: A Flexible and Efficient Split Federated Learning Framework for Edge Heterogeneity.

Area of Science:

• Edge Computing
• Machine Learning
• Distributed Systems

Background:

• Federated Learning (FL) deployment in edge environments is hindered by system heterogeneity, non-IID data, and communication constraints.
• These factors impede training efficiency and scalability in edge AI systems.

Purpose of the Study:

• To present FLEX-SFL, a flexible and efficient split federated learning framework.
• To jointly optimize model partitioning, client selection, and communication scheduling for edge environments.

Main Methods:

• Device-aware adaptive segmentation strategy to mitigate straggler effects.
• Entropy-driven client selection algorithm for data representativeness.
• Hierarchical local asynchronous aggregation for improved throughput and reduced latency.

Main Results:

• FLEX-SFL demonstrates superior model accuracy, convergence speed, and resource efficiency compared to state-of-the-art baselines.
• Experiments on FMNIST, CIFAR-10, and CIFAR-100 datasets validate performance under high heterogeneity.
• Theoretical convergence properties established under convex settings.

Conclusions:

• FLEX-SFL effectively addresses challenges in edge FL, enhancing training efficiency and scalability.
• The framework's coordinated mechanisms prove practical for real-world edge intelligent systems.