Energy-Efficient Federated Transfer Learning in 6G Native AI Networks

In 6G, the proliferation of artificial intelligence (AI) applications and the collaboration of multi-agents require ubiquitous computing, communication, intelligence and security on wireless networks. However, traditional data-driven learning approaches require an ample amount of original data and trains AI models centrally, which not only significantly aggravate privacy concerns, but also burden the resource-limited terminal devices. Federated learning and transfer learning are expected to deal with these challenges through distributed training paradigms and knowledge sharing among different learning algorithms. In this paper, a novel federated transfer learning framework is proposed for 6G native AI, which trains local models on distributed edge multi-agents and performs global aggregation on edge base station. Energy consumption is minimized by jointly optimizing neural network and power allocation. Then a deep deterministic policy gradient based algorithm is proposed to address the continuous action space with polynomial complexity. Simulation results illustrate that the proposed algorithm can significantly enhance model accuracy and decrease energy consumption compared with benchmarks.