Performance Difference Based Lazy Aggregation in Federated Learning Human Activity Recognition

With the popularity of the concept of smart life and the rapid development of wearable terminal device technology, human activity recognition based on sensor data has gained widespread attention. To avoid the data privacy and security issues brought about by the centralized training model, federated learning is introduced; however, aggregation of local models requires a large amount of information to be transferred between the client and the server, consuming the network bandwidth, which is huge for some terminal devices. It may impede the final aggregation of the models. In this paper, we aim to save communication consumption by reducing the communication from the client to the server, the core idea of which is to reuse previous updates to skip less informative communication. We conducted experiments on the WISDM and UNIMIB SHAR datasets, and the results show that our method can reduce up to 56.8% and 37.5% of communication on the WISDM and UNIMIB SHAR datasets, respectively, without any significant performance degradation compared to the FedAvg algorithm. At the same time, our method can synergize with existing quantization algorithms to further reduce communication consumption, rather than replacing existing methods.