CLAP: Cooperative learning through partial model exchange for heterogeneous wireless sensor networks
Joannes Sam Mertens, Laura Galluccio, Giacomo Morabito · Internet of Things · 2025
In the recent past there has been an increasing interest in the execution of Machine Learning (ML) in resource-constrained devices, for example in wireless sensor networks (WSNs). In such a context, the development of cooperative learning protocols has gained considerable attention for its privacy-preserving and communication-efficient capabilities. However, the application of such cooperative learning techniques presents some criticalities in wireless sensor networks, especially when there is diversity in ML tasks to be executed or in the type of data collected by sensors. In this paper, we propose a cooperative learning algorithm called CLAP that partitions the neural network into layers specific of the type of data collected by the sensors and layers specific of the ML task executed. Furthermore, in CLAP a WSN setting is considered where certain nodes can perform both training and inference, while others can only perform inference, due to intrinsic differences in hardware and processing capabilities. To take such heterogeneity into account, CLAP employs a clustering strategy in which Cluster Heads are nodes with large computing resources and long-range communication radios, whereas Cluster Members are low capability nodes with short-range radios. CLAP performs a cooperative learning algorithm where only parts of the ML model parameters are exchanged, so improving communication efficiency and privacy. Numerical results assess the effectiveness of the proposed approach.