Hardware setup

This project’s hardware setup is partitioned into three core components: Clients 1 and 2 and a central server, as depicted in Fig.1. Each of these segments is equipped with essential hardware, including NI USRP 2900 and Xbee S2C devices, as shown in Fig.3. The NI USRP 2900, mounted with an omnidirectional VERT 2450 antenna, is utilized for physical layer communication, while Xbee S2C devices manage network-level synchronization. To enhance system robustness and eliminate the need for channel estimation, the USRPs leverage differential modulation techniques.

The project also incorporates key signal specifications that are crucial for the communication system’s functionality. These specifications include an operating frequency of 2.5 GHz, pulse shaping using Root Raised Cosine, a reception gain of 20 dB, an IQ rate of 50 kSps, a symbol rate of 6.25 kSps with 8 samples per symbol, and a default transmission power of 20 dB.

The video below showcases a real-time demonstration of our project, highlighting the setup and operation of the federated learning environment using NI USRP 2900 and ZigBee S2C devices. In this demo, we explain how the hardware components are integrated to facilitate seamless communication of model updates across a bandwidth-constrained wireless channel. The video walks through the entire process, from local model training at the client devices to the transmission of updates to the central server, ultimately resulting in a generalized global model. This real-time demo provides a clear visualization of the practical implementation of our prototype and demonstrates the effectiveness of our configuration in achieving decentralized, privacy-preserving machine learning.