An efficient Resource Allocation framework using federated LSTM for Network Function Virtualization

Abstract

Network Function Virtualization (NFV), a new technology emerging within the telecommunications space, get lower operating and capital costs while enabling service deployment. By co-locating various types of network appliances on standard, high-volume servers, switches and storage across the industry, NFV provides a means to resolve challenges associated with network services, by using standard information technology (IT) virtualization technologies. In addition, researchers are still continuing to solve other challenges of NFV, these research challenges include managing and orchestrating Virtual Network Functions (VNFs), chaining services through service chain (SFC), scheduling VNFs for low latency with minimum overhead, and providing efficient allocation of virtual network resources/functions within the Network Function Virtualization Infrastructure (NFVI) to create a flexible and dynamic network that supports multiple demands. Another significant challenge in the NFV space is the inability to support the current levels of inefficiency, SFC complexity, increased latency, load imbalance and poor scalability prevalent today in NFV systems. The paper proposes a new intelligent NFV resource allocation framework based on Long Short-Term Memory (LSTM) networks with Federated Deep Reinforcement Learning (FDRL) to overcome the above-mentioned challenges. The process of the proposed solution consists of multiple phases as follows: VNF placement and resource are scaling using Deep Reinforcement Learning (DRL) to facilitate dynamic and adaptive NFV management and executing (performing) the proposed intelligent NFV resource allocation solution through implementation of the above-mentioned DRL Federated Learning (FL) can be done through distributed model training in order to provide privacy-preserving model optimization without directly sharing model data and LSTM-based traffic prediction, which provides predictions of future traffic demand based upon a historic pattern of NFV resources used. By using both methods together to solve this problem, we expect to create a more efficient network, with reduced latency, and improved usage of available resources.