Neuron Pruning-Based Federated Learning for Communication-Efficient Distributed Training

Efficient and flexible cloud computing is widely used in distributed systems. However, in the Internet of Things (IoT) environment with heterogeneous capabilities, the performance of cloud computing may decline due to limited communication resources. As located closer to the end, edge computing is used to replace cloud computing to provide timely and stable services. To accomplish distributed system and privacy preserving, Federated Learning (FL) has been combined with edge computing. However, due to the large number of clients, the amount of data transmitted will also grow exponentially. How to reduce the communication overhead in FL is still a big problem. As a major method to reduce the communication overhead, compressing the transmission parameters can effectively reduce the communication overhead. However, the existing methods do not consider the possible internal relationship between neurons. In this paper, we propose Neuron Pruning-Based FL for communication-efficient distributed training, which is a model pruning method to compress model parameters transmitted in FL. In contrast to the previous methods, we use dimensionality reduction method as the importance factor of neurons, and take advantage of the correlation between them to carry out model pruning. Our analysis results show that NPBFL can reduce communication overhead while maintaining classification accuracy.