Federated learning (FL) is a field in distributed optimization. Therein, data collection and neural network (NN) training are decentralized, which means these tasks take place in many clients with limited communication and computation capabilities. In FL, the client NNs are trained with locally available data. Next, the client NNs are aggregated in order to update a global NN.

In this work, we apply FL on a small live gesture sensing NN for low-power 60GHz frequency modulated continuous wave (FMCW) radar from Infineon Technologies. Furthermore, we investigate the effects of client data heterogeneity on the gesture recognition accuracy. The presented FL algorithms are evaluated on a diverse dataset including approximately 26k gesture recordings. Each recording contains a specific gesture execution sequence, which is labeled accordingly as that distinct gesture. The rest of the recording, which does not contain any specific gesture, is simply labeled as the background. To study FL with varying data heterogeneity, the data among different clients is partitioned in two different ways. The first partition, called independently and identically distributed (iid) partition, involves shuffling and equally distributing the dataset among clients using the same number of local epochs during training. The second partition is non-iid, where each client is assigned a different number of gesture recordings and local epochs. It is shown that, FL converges in the iid partition to an accuracy higher than 92.4%. However, the increasing data heterogeneity degrades the accuracy to 78.3%. To address accuracy degradation resulting from client heterogeneity, we propose dynamically weighting the recordings during training based on the varying ratio of distinct gesture sequences and background in each client. Moreover, regularization terms are included in the loss function to prevent client drift and overconfidence in the local NN predictions. Finally, it is shown that the proposed adaptions reduce the accuracy degradation, such that 96.4% label accuracy, with the highest degree of data heterogeneity (one gesture per client), is obtained.