A federated learning out-of-distribution generalization method based on causal inference

Abstract

The application is a federated learning out-of-distribution generalization method based on causal inference, and belongs to the technical field of edge intelligent devices. The method comprises the following steps: constructing a federated learning framework to train a task model, in each round of training, a server sends a global task model parameter and a global prototype library to a client participating in training; the client trains a local model by using a local sample, obtains individualized features and global features of the sample by a decoupling representation learning module, and obtains causal features after removing pseudo-association by a mixed factor removing training module; the client sends updated model parameters and various local prototypes to the server; the server updates the global prototype library and the global model parameter by using a hierarchical update strategy module; the training is repeated until a convergence condition is met, and the obtained global task model is deployed to a new device for use. The application improves the generalization ability of the task model when facing invisible distribution data, and ensures the usability and applicability of the task model.

Images