VNF Placement based on Resource Usage Prediction using Federated Deep Learning Techniques

In recent years, virtual network functions (VNFs) and multi-cloud environments have emerged that allow virtual services to be hosted dynamically across multiple clouds with greater efficiency and cost optimization. VNFs require dynamic instantiation and resource allocation to handle the dynamic traffic demands. The purpose of this paper is to propose a deep learning technique for predicting the resource requirements of VNFs in order to efficiently identify locations to deploy VNFs, improve resource utilization and minimize over- and underallocations. In addition, VNFs deployed in multi-cloud provider environments present a serious privacy risk because cloud providers may not provide access to the VNF usage data to another cloud provider which may be needed for more accurate learning. This issue can be addressed by training deep learning models using federated learning (FL), where the data remains with each cloud provider, but a shared model is learned globally. The trained models can then be integrated into any standard VNF placement algorithm. In this work, we evaluate the performance of our proposed FL approach for VN resource prediction using synthetic and real-world data. The FL model is compared with a centralized deep learning model and other standard machine learning models. In our evaluation, we noticed that appropriate deep learning models helped prevent resource over- and under-allocation by enhanced prediction of resource requirements, resulting in lower resource consumption and better service quality.