Hierarchical Optimization for Asynchronous Vertical Federated Learning

Vertical federated learning (VFL) is a privacy-preserving computing method that models heterogeneous data from multiple parties to train a better predictor. However, the traditional ways of VFL model training are mainly based on synchronous computing. Due to differences in each party's feature sizes, computation resources, and network conditions, one party often waits for the other. To improve the efficiency of VFL, we have proposed a new asynchronous VFL training framework. Accordingly, we have split the VFL model into the client and aggregator models, and then use hierarchical optimization to asynchronously update the models of the client and aggregator, allowing the client to save time waiting for the aggregator or other clients. To ensure that the client entirely obtains task information during training, each party also builds an auxiliary supervision classifier based on its client model to fit labels directly. Empirically, we apply our method to the vertical federated neural network model. The test results on several datasets show that our approach improves the efficiency of model training and reaches an accuracy comparable to synchronous methods.