Communication-Efficient Cluster Federated Learning in Large-scale Peer-to-Peer Networks
share
A traditional federated learning (FL) allows clients to collaboratively train a global model under the coordination of a central server, which sparks great interests in exploiting the private data distributed on clients. However, once the central server suffers from a single point of failure, it will lead to system crash. In addition, FL usually involves a large number of clients, which requires expensive communication costs. These challenges inspire a communication-efficient design of decentralized FL. In this paper, we propose an efficient and privacy-preserving global model training protocol in the context of FL in large-scale peer-to-peer networks, CFL. The proposed CFL protocol aggregates local contributions hierarchically by a cluster-based aggregation mode, as well as a leverged authenticated encryption scheme to ensure the security communication, whose key is distributed by a modified secure communication key establishment protocol. Theoretical analyses show that CFL guarantees the privacy of local model update parameters, as well as integrity and authenticity under the widespread internal semi-honest and external malicious threat models. In particular, the proposed key revocation based on public voting can effectively defense against external adversaries hijacking honest participants to ensure the confidentiality of the communication keys. Moreover, the modified secure communication key establishment protocol indeed achieves high network connectivity probability to ensure transmission security of the system.
READ FULL TEXT