Abstract:

Federated learning is an effective method to solve the problem of multi-organization collaborative training. However, existing federated learning has problems such as not supporting user disconnection and model API leaking sensitive information. This paper proposes a user-oriented federated learning data privacy preserving method that supports user disconnection, which can train a differential privacy disturbance model under user disconnection and protected model parameters. This paper uses a federated learning framework to design a data privacy preserving model based on deep learning. It mainly contains two execution protocols, server and user execution protocol. User trains a deep model locally, adds differential privacy disturbance to the local model parameters, and adds sum noise of dropped users to the aggregated parameters so that the federated learning process meets (ε,δ)-differential privacy. Experiments show that when the number of users is 50 and ε=1, a balance can be reached between model privacy and usability.

Key words: federated learning, deep learning, privacy preserving, differential privacy, user disconnection