Personnel status detection model suitable for vertical federated learning structure
Jie Ji, Danfeng Yan, Zhengyang Mu · 2022
With the improvement of the medical system, the universal access of wearable devices and people's greater concern about personal health, personnel health detection has received greater attention. However, existing personnel status detection faces the problem that existing models cannot get more accurate results by multiple types of data because the data is difficult to obtain across domains due to privacy issue. In this paper, we propose a vertical federated learning model for person status detection using multimodal data. Considering the varying compute capability of training devices and the different inference speed caused by modality complexity, a lightweight feature extraction sub-model is used, while ensuring data privacy and tolerating accuracy degradation. Also, we design a Fast and Secure (FS) module that effectively reduces the amount of data onto the network transmission process, the model's dependence on the network and the risk of data being intercepted by the third party. Experiment results demonstrate the feasibility of multimodal vertical federated learning and the accuracy is increased by 5.6 percentage points compared with unimodal. The FS module reduces the amount of network transmission information by 50% and 48.1% during model use and model training, respectively