Federated split learning for sequential data in satellite-terrestrial integrated networks

Satellite-terrestrial integrated networks (STINs) have been proposed for B5G/6G mobile communication, and the increase in the computation and communication capacities of satellites opens the door for satellite edge computing in space. The requirements of privacy protection and communication efficiency lead to the inefficiency or deficiency of traditional learning approaches in STINs. This study incorporates federated learning and split learning paradigms with STINs and introduces a split-then-federated learning framework and federated split learning with long short-term memory to handle sequential data in STINs. A case study of electricity theft detection based on a real-world sequential power load dataset is utilized to demonstrate the effectiveness of the proposed solution. Numerical experiments show that the proposed FedSL-LSTM model is an effective federated learning solution and a competitive solution compared with centralized models in terms of classification performance evaluation metrics.

Automated summary

This study incorporates federated learning and split learning paradigms with satellite-terrestrial integrated networks and introduces a split-then-federated learning framework and federated split learning with long short-term memory to handle sequential data in STINs. The proposed solution is demonstrated to be effective through a case study of electricity theft detection based on a real-world dataset.