Security defense method for data manipulation attacks in federated learning

Abstract

The invention discloses a security defense method for data manipulation attacks in federated learning. The method comprises the following steps: 1, enabling a central server to receive local model parameters uploaded by users, and calculating the similarity degree of the local model parameters uploaded by each user and a fusion coefficient corresponding to each user, wherein the local model parameters are local model parameters obtained after the users adopt private training data to train for one round; 2, after the central server receives the local model parameters of the local users in one period, enabling the central server to calculate a weighted average value of the local model parameters of each user according to the fusion coefficient to obtain global model parameters, wherein the period is a preset local model parameter updating round number; 3, issuing the global model parameters to the corresponding users, and enabling the users to update the local model parameters after receiving the global model parameters.

Description

technical field [0001] The present invention relates to a method for realizing a security training framework of federated learning, and more specifically relates to a method for realizing a security training framework for defending against data manipulation attacks in federated learning. Background technique [0002] Federated learning is an emerging deep learning framework. In the traditional centralized deep learning, the central server needs to collect a large amount of user data for training the neural network model (referred to as the model), but due to the large network communication overhead of data transmission, the ownership of user data and the privacy of user data, etc. , user data for deep learning is often difficult to obtain. And federated learning adopts another way of training the neural network model: in a round of training, each user uses its private data to train the local model, and then uploads the parameters of the local model to the central server, an...

AI Technical Summary

Problems solved by technology

However, existing research work lacks an analysis of how to defend against threats posed by users participating in federated learning

Since the threat of data manipulation attacks comes from within users participating in federated learning, malicious users can hide among normal users, their attack methods are more hidden, and the attack methods are more difficult to detect and troubleshoot. Existing methods for security issues in federated learning It is difficult for existing defense technologies to effectively resist the above two data manipulation attacks, which also brings huge security risks to users participating in federated learning

Images