Learning from Peers at the Wireless Edge

The last mile connection is dominated by wireless links where heterogeneous nodes share the limited and already crowded electromagnetic spectrum. Current contention based decentralized wireless access system is reactive in nature to mitigate the interference. In this paper, we propose to use neural networks to learn and predict spectrum availability in a collaborative manner such that its availability can be predicted with a high accuracy to maximize wireless access and minimize interference between simultaneous links. Edge nodes have a wide range of sensing and computation capabilities, while often using different operator networks, who might be reluctant to share their models. Hence, we introduce a peer to peer Federated Learning model, where a local model is trained based on the sensing results of each node and shared among its peers to create a global model. The need for a base station or access point to act as centralized parameter server is replaced by empowering the edge nodes as aggregators of the local models and minimizing the communication overhead for model transmission. We generate wireless channel access data, which is used to train the local models. Simulation results for both local and global models show over 95% accuracy in predicting channel opportunities in various network topology.