Adaptation of Multi-Exit Architecture for Efficient Federated Learning

This paper explores the integration of a multi-exit architecture into federated learning (FL) to address the challenges posed by varying computational resources and non-uniform data distributions across client devices. By introducing multiple output layers, the multi-exit architecture allows clients to perform early exits based on their resource capacity, reducing computational load and communication costs. Experimental results show that this approach enhances model accuracy by up to 8.15% and reduces communication costs by 15.4%, making it a viable solution for improving FL efficiency in heterogeneous environments.