Abstract

The rise of intelligent edge services has intensified the demand for scalable and privacy-preserving collaborative learning frameworks. Split Federated Learning (SFL), a hybrid paradigm that combines the layer-wise decoupling of Split Learning with the distributed aggregation of Federated Learning, offers an efficient training approach across heterogeneous devices. However, its scalability remains constrained by device heterogeneity and synchronization delays. This thesis proposes a novel Collaborative Split Federated Learning (CSFL) framework that facilitates real-time cooperative computation through direct device-to-device communication. At its core lies the Collaborative Relay Optimization Mechanism (CROM), in which efficient devices act as computational relays by executing intermediate model segments on behalf of bottleneck devices, thereby achieving balanced workload distribution. To minimize end-to-end training latency, we formulate a joint optimization problem over the model cut-layer and device pairing configuration. This problem is decomposed into two interdependent sub-tasks: cut-layer selection, addressed via a convergent alternating optimization strategy, and device pairing, modeled as a deferred-acceptance-based weighted matching game that aligns local utilities with global objectives. Extensive experiments on VGG-16 using the Tiny-ImageNet dataset under non-IID data partitions demonstrate that CSFL significantly reduces training latency and energy consumption while maintaining accuracy comparable to conventional SFL. Overall, this work advances collaborative edge learning by introducing a unified framework for system-level co-optimization, emphasizing the importance of algorithm-architecture codesign for future intelligent systems.