QuantFL: Pre-Training-Aware Update Quantisation for Communication-Efficient Federated Learning in the 6G IoT-Edge-Cloud Continuum

Abstract

Future 6G IoT-edge-cloud systems will rely on distributed intelligence, where edge devices adapt shared models using local private data. Federated learning (FL) supports this setting, but repeated transmission of full-precision model updates remains expensive over bandwidth-limited and energy-constrained wireless links. This paper presents QuantFL, a communication-efficient FL framework that starts from a pretrained initialisation and compresses per-layer client updates using bucket-based scalar quantisation. The key observation is that pre-training reduces the dynamic range and variance of local updates, making simple scalar quantisation more effective at a fixed number of code levels. Clients transmit compact bucket indices every round, while codebooks, defined as bucket boundary lists, are refreshed infrequently and amortised in the communication budget. We evaluate Bucket-Uniform (BU) and Bucket-Quantile (BQ) quantisers against QSGD and non-quantised FedAvg on MNIST and CIFAR-100 under non-IID data. BU provides the strongest accuracy-communication tradeoff, achieving competitive accuracy with approximately 40% total uplink+downlink reduction when the downlink remains full precision and substantially larger savings on the uplink, which is often the bottleneck in wireless edge learning. We further discuss compatibility with error feedback and position QuantFL as a lightweight update-compression primitive for scalable edge intelligence.