Copyright (C) <2021-> by Mobile Systems and Networking Group, School of Computer Science, Fudan University

Nowadays, a large number of users are using wearable devices, such as smartwatches, which can only be used to collect limited user information features. Under this limitation, federation learning with MLP model is unable to achieve good performance in classification tasks. To tackle this problem, we design a classification model based on LSTM and attention mechanism to learn the temporal relationship in limited information and achieve a better Human Activity Recognition (HAR) performance. To deal with the non-IID distribution problem of labels indifferent group, we involve the federated learning method FedRS. We found that in the federated learning task for wearable devices, when the main parameters of FedRS α is configured as a higher value, the model can be trained to reach a better performance. Differently, a smaller α will make local training difficult to converge. Cloudlets are used to reduce the local training load for wearable devices, and thus save the battery life of them. The use of cloudlets could also reduce the communication cost of different groups in the federated learning framework.

• Python
• PyTorch

The first dataset is Daily and Sports Activities Data Set (DSA), mainly the motion sensor data of 19 daily human activity traces. The second dataset is Physical Activity Monitoring for Aging People (PAMAP), containing signals of 8 participants performing 14 activities.

The designed model Att-LSTM is composed of one LSTM layer, one attention layer, and one output layer. We adopt FedRS to limit the update of missing class weights during the local training. The baseline models are implemented by MLP, composed of one input layer and one output layer, also one hidden layer with 1000 units using ReLU activations and linear SVM.

• Paper: "CloudyFL: A Cloudlet-Based Federated Learning Framework for Sensing User Behavior Using Wearable Devices"

@article{gong2022cloudyfl,
  title={CloudyFL: A Cloudlet-Based Federated Learning Framework for Sensing User Behavior Using Wearable Devices},
  author={Gong, Qingyuan and Ruan, Hui and Chen, Yang and Su, Xiang},
  booktitle = {Proceedings of the 6th International Workshop on Embedded and Mobile Deep Learning},
  year={2022}
}

[1] Billur Barshan and Murat Cihan Yüksek. 2013. Recognizing Daily and Sports Activities in Two Open Source Machine Learning Environments Using Body-Worn Sensor Units. The Computer Journal 57, 11 (2013), 1649–1667.

[2] Xin-Chun Li and De-Chuan Zhan. 2021. FedRS: Federated Learning with Restricted Softmax for Label Distribution Non-IID Data. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining.

[3] Attila Reiss and Didier Stricker. 2011. Towards global aerobic activ- ity monitoring. In Proceedings of the 4th International Conference on PErvasive Technologies. ACM, 12.