Two Birds With One Stone: Towards Communication and Computation Efficient Federated Learning

Federated learning (FL) is a distributed machine learning paradigm proposed to ensure user data privacy. However, in practical scenarios, terminal devices often contend for limited network resources, making communication overhead a significant obstacle to its real-world application. Existing algorithms typically trade computational overhead to reduce communication overhead, while also relying on auxiliary datasets. In consideration of real-world scenarios, we propose a federated learning framework based on autoencoders and the attention mechanism. This framework guarantees training performance and low communication overhead while reducing the computational overhead on the server for parameter reconstruction. Experimental validation on multiple datasets demonstrates the algorithm’s performance.