Covert Hybrid Federated and Centralized Learning

This work addresses latency minimization in hybrid federated and centralized learning (HFCL) systems while ensuring covert communication (CC) in the presence of an eavesdropper using a friendly jammer. An optimization problem jointly optimizes transmission power, jammer zero-forcing precoding, and uplink delays for clients while ensuring the covertness of legitimate links. Furthermore, useful insights into the achievable covertness level with the proposed scheme are presented. Simulation results highlight trade-offs between latency and system parameters, including the covertness level, bandwidth, jamming power, number of clients performing centralized learning, and eavesdropper location. Valuable insights can be gained regarding application enabling criteria such as accuracy and latency, while the proposed scheme ensures optimal operation of HFCL systems when efficient and secure communication is required.