FedCPD: A Federated Learning Algorithm for Processing and Securing Distributed Heterogeneous Data in the Metaverse

Abstract

The continuous development of virtual reality technology allows the metaverse to create more immersive and highly interactive experiences for users. Metaverse users upload personal information through virtual reality devices, causing data security and communication security issues. Moreover, the diversity of data sources within the metaverse exacerbates issues of data heterogeneity. To address these issues, we propose a generative learning-based federated learning algorithm to secure and process heterogeneous data from users in the metaverse, called FedCPD. It consists of three main modules: a privacy protection module for data security, a correction module to correct the bias of the classifier, and an aggregation module to improve model performance. To protect the data security of metaverse users, we design a privacy-preserving method based on conditional Generative Adversarial Networks (cGAN) in the privacy protection module. The method replaces the feature extractor with a generator in cGAN to engage in server-side aggregation to avoid data exposure. The correction module is proposed to enhance the classifier’s ability to classify unknown data by using the constructed pseudo dataset for classification model training. To alleviate the negative impact of data heterogeneity on the global model, the aggregation module utilizes local discrepancy-based aggregation weights for server-side aggregation. It assigns higher aggregation weights to the client models that perform better than other models. Extensive experiments on multiple datasets show that FedCPD exhibits the highest classification accuracy compared to existing algorithms, demonstrating its effectiveness in processing heterogeneous data.

Keywords

• Metaverse
• federated learning
• generative learning
• data security
• data heterogeneity