Efficient lower layers parameter decoupling personalized federated learning method of facial expression recognition for home care robots

Facial expression recognition (FER) is a crucial and pivotal functionality for home care robots that engage in intimate interactions with human individuals. However, the potential privacy risk associated with home care robots resides in their acquisition of facial information during expression recognition and subsequent data uploading for model updates. Federated learning (FL) can safeguard data privacy while enabling machine learning, yet its application to home care robots poses several challenges, including non -independent and non -identically distributed collected data, limited communication and computation capabilities on the robot side. However, limited research has been conducted in this area. In this paper, we propose a lightweight neural network architecture for the federated FER model on home care robots. We also introduce an efficient and simple lower -layer personalized FL method that optimizes local personalized models by exchanging fewer parameters. Furthermore, we implement a demo on real physical robots NAO to demonstrate how our federated model improves the performance of local FER models in non-IID data settings. The results demonstrate that our approach achieves competitive performance on four simulated non-IID facial expression recognition (FER) datasets, comparable to larger network models and state-of-the-art algorithms. There is an average difference of approximately 3% from the best algorithm results on three datasets, while outperforming the best algorithm by 2% on one dataset. The evaluation of our methods on two real NAO robots shows that the overall performance improvement of our federated model over the locally independent model is about 20% on average. The open -source code is available at https://github.com/zxecho/FESAHR.