A Federated-CNN based Proactive Caching Algorithm for vCDN System

In order to improve the speed of user's access to the contents on the Internet and reduce the flow into the core network, the virtual content delivery network (vCDN) which can support flexible deployment is proposed as a solution. A crucial problem in vCDN system is to improve the hit rate of cache on the premise of reducing the burden of vCDN nodes. At present, the main method to improve the cache hit rate is using proactive cache algorithm based on popularity prediction. There have been researches on proactive caching algorithms, such as LSTM, Random Forest and KNN. But when the training data is insufficient, the prediction performance of these models is not good enough. If collecting data from distributed nodes to improve training performance, it will bring a lot of communication overhead and cannot well protect user privacy. These problems are more notable in vCDN system because of its high dispersion. In order to solve the above problems, we propose a proactive cache algorithm based on federated learning (FL) framework. Distributed vCDN nodes use their own data to train convolutional neural network (CNN) model to predict the popularity of future content. Then, through the federal average of each node model, the model for predicting the global popularity can be obtained. Each vCDN node uses the global model to predict the incoming requests, which can improve the caching efficiency of the node by predicting popularity. Federated-CNN complete the training locally instead of centralizing all the user's data. This can reduce transmission and protect the user privacy at the same time. The simulation results show that through federated aggregation, the accuracy of Federated-CNN in popularity prediction is about 22 % higher than that of other proactive caching algorithms, and by using FL, the training efficiency increases over 10 times.