An Efficient Content Popularity Prediction of Privacy Preserving Based on Federated Learning and Wasserstein GAN

Abstract

To relieve the high backhaul load and long transmission time caused by the huge mobile data traffic, caching devices are deployed at the edge of mobile networks. The key to efficient caching is to predict the content popularity accurately while touching the users’ privacy as little as possible. Recently, many studies have applied federated learning in content caching to improve data security. However, they still give away some privacy of participants, especially ignoring the private data leakage in the trained model. To solve this problem and further improve the cache hit ratio, we propose an efficient content popularity prediction of privacy-preserving (CPPPP) scheme based on federated learning and Wasserstein generative adversarial network (WGAN), which achieves a high cache hit ratio. Benefited by the Federated-WGAN and the generated fake samples, the private data, the content preferences of individual users, and so on are well protected. In particular, gradient clipping and model parameter limitation are introduced in the model training, and the security of the modified model is greatly improved compared with the original model. Results show that the proposed scheme has a higher cache hit ratio than the existing federated learning-based methods while limiting the privacy leakage caused by the trained model to a quite low level.