Backdoor Defense Method of Vertical Federated Learning Based on Clustering of Neuron Activation Values

Abstract

The invention discloses a vertical federated learning backdoor defense method based on neuron activation value clustering, including: constructing a vertical federated recommendation system including multiple participants and collaborators; After the product-oriented links are constructed, the product-oriented links of potential backdoor attacks are effectively screened out by classifying the product-oriented links, and the product-oriented links of backdoor attacks are repaired by using the clustering results to guide the direction of backdoor attacks. Commodity samples are learned towards the correct prediction direction, so that the repaired commodity recommendation model can defend against backdoor attacks without obtaining the commodity samples of the participants; it also filters out the joint embedding representation with the same ID as the aggregate embedding representation of the backdoor attack Or attack repair, to prevent or improve the aggregation embedding representation of backdoor attacks, optimize the parameters of the product recommendation model, and improve the defense ability of the product recommendation model against backdoor attacks.

Description

technical field [0001] The invention belongs to the technical field of privacy security for product recommendation, and in particular relates to a backdoor defense method based on neuron activation value clustering for longitudinal federated learning. Background technique [0002] Deep learning has achieved excellent performance in complex and variable recommendation tasks, which makes it widely used and deployed in commodity recommendation systems in real scenarios. Compared with the traditional product recommendation system, the product recommendation system built by deep learning has further improved performance, which is mainly due to abundant computing resources and sufficient computing data. However, in recent years, some countries and regions have introduced data privacy protection regulations that restrict the large-scale collection of commercial data, which brings great challenges to the product recommendation system built by deep learning, and the following is the ...

AI Technical Summary

Problems solved by technology

However, these two technologies are not suitable for product recommendation systems to defend against the threat of backdoor attacks. The differential privacy technology makes the accuracy of product recommendation tasks drop sharply by adding random noise, and cannot balance the performance of backdoor defense and recommendation task performance.

Gradient sparseness cannot defend the active party of the vertical federated recommendation system from backdoor attacks, mainly because the active party launches backdoor attacks that do not depend on the gradient information of the model

Images