A gradient perturbation-based federated learning data privacy protection method and system

Abstract

The invention discloses a federated learning data privacy protection method and system based on gradient perturbation, which belongs to the field of data privacy protection. The method includes: using a local model trained by federated learning to perform category prediction on samples in data participants to obtain the original prediction Probability vector; the original predicted probability vector is perturbed to obtain a perturbed prediction whose predicted label is the same as that of the original predicted probability vector and whose gradient of the predicted loss function has the largest angular deviation from the gradient of the predicted loss function of the original predicted probability vector Probability vector; with the goal of minimizing the difference between the original prediction probability vector and the disturbance prediction probability vector of each local model, retrain each local model; aggregate the retrained local models to obtain a global model. The protected federated learning global model can effectively reduce the risk of model prediction output and model gradient leaking the privacy of user participants while maintaining the availability of the model.

Description

Technical field [0001] The present invention belongs to the field of data privacy protection, and more particularly to a federal learning data privacy protection method and system based on gradient disturbance. Background technique [0002] Depth learning as a method of realizing artificial intelligence, has been applied to many fields such as computer vision, data mining and medical diagnosis. Depth study requires massive training data as support, making it challenges for data security and privacy protection. In addition, if you are worried about data privacy, each data participant refuses to share its data information, and the central server cannot be trained to get an efficient and reliable learning model to form data island. The increasingly stringent data privacy protection requirements and data is island have seriously restricted the development and application of deep learning. In order to solve these problems, federal learning came into being. However, for federal learnin...

AI Technical Summary

Problems solved by technology

However, under high privacy protection performance requirements, these defense schemes will introduce too much noise perturbation, which will greatly reduce the usability of the model

Images