Consistent Vertical Federated Deep Learning Using Task-Driven Features to Construct Integrated IoT Services

Abstract

By training a multivariate deep learning model distributed across existing IoT services using vertical federated learning, expanded services could be constructed cost-effectively while preserving the independent data architecture of each service. Previously, we proposed a design approach for vertical federated learning considering IoT domain characteristics. Also, our previous method, designed leveraging our approach, achieved improved performance, especially in IoT domains, compared to other representative vertical federated learning mechanisms. However, our previous method was difficult to apply in real-world scenarios because its mechanism consisted of several options. In this paper, we propose a new vertical federated learning method, TT-VFDL-ST (Task-driven Transferred Vertical Federated Deep Learning using Self-Transfer partial training), a consistent single mechanism even in various real-world scenarios. The proposed method is also designed based on our previous design approach. However, the difference is that it leverages a newly proposed self-transfer partial training mechanism. The self-transfer partial training mechanism improved the MSE and accuracy of TT-VFDL-ST by 0.00262 and 12.08% on average compared to existing mechanisms. In addition, MSE and accuracy improved by up to 0.00290 and 5.08% compared to various options of our previous method. By applying the self-transfer partial training mechanism, TT-VFDL-ST could be used as a key solution to construct real-world-integrated IoT services.