An Unequal Error Protection-based Coded Transmission for Federated Learning

Abstract

Effective data transmission plays a crucial role in federated learning (FL), which enables collaborative model training without centralizing data. This paper proposes a new coded transmission to enhance the communication quality for FL. The proposed coded transmission incorporates weight quantization, multilevel coding, set partitioning, and multi-stage decoding which are optimized to improve the FL performance. Furthermore, the unequal error protection (UEP) strategy is adopted in the proposed coded transmission, which allows the code rates to be optimized according to the significance of the quantized data. Simulation results demonstrate that the proposed UEP-based coded transmission outperforms conventional bit-interleaved coded modulation (BICM) scheme in terms of NMSE performance for FL, which, in return, improves the FL performance.