Dynamic weight quantization for embedded federated learning in LoRa mesh networks: A case study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The integration of federated learning (FL) into resource-constrained devices promises a re-trainable distributed Machine Learning (ML) intelligence in Internet of Things (IoT) networks. However, in low-power wide-area network (LPWAN) technologies such as LoRa, the extremely limited and dynamically varying bandwidth at each node poses a major obstacle to the practical deployment of FL. While static model quantization has been proposed to reduce communication overhead, it cannot adapt to changing network conditions. This paper investigates dynamic weight quantization (DynQuant) as a communication-aware strategy for embedded federated learning over LoRa networks. DynQuant enables both clients and the server to flexibly select the quantization bit width at each FL round based on dynamic system and network conditions. We first explore the design space of dynamic quantization through extensive simulations, analyzing the tradeoff between model accuracy, quantization strategy, and communication cost. Subsequently, we validate the practical feasibility of the approach through a prototype implementation on real microcontroller-based nodes interconnected via a LoRa mesh network. Our results show that dynamic quantization can significantly reduce the volume of transmitted data and communication time, while only slightly affecting the final model accuracy. For instance, early quantization achieves an 8.6% reduction of communicated model data and the asymmetric 6 → 8 quantization strategy achieves a 12.5% reduction in communicated model data compared to static 8-bit quantization. These findings support dynamic quantization as a practical and effective mechanism for tailoring federated learning to dynamic, low-data-rate IoT networks.