A neural network-based vertical federated learning framework with server integration

Federated learning trains an algorithm from data stored separately at the clients without exchanging the raw data, preserving privacy. In a vertically partitioned federated learning setup, for the same data record there are input features uniquely located at a subset of clients. Most existing works on vertical federated learning support only two participating clients and assume that each client has the output feature. In this work, a vertical federated learning framework based on the neural network is proposed which can be employed with any number of clients that also incorporates a server as a third party and it is assumed that only one client has the output feature. The convergence rate and the accuracy remain the same as when the data is centralized. Privacy analyses demonstrate that the exchange of hidden layers’ outputs among the participating clients does not cause any data leakage nor can any participant extract any valuable information from the other. The proposed vertical federated learning framework exhibits its potential applications in engineering by facilitating secure and collaborative model training across distributed datasets, presenting promising opportunities for improved predictive models and optimization of critical infrastructure systems.