| Title | : | Federated Learning approach for Auto-scaling of Virtual Network Function resource allocation in 5G-and-Beyond Networks |
| Speaker | : | Rahul Verma (MS Scholar - IITM) |
| Details | : | Tue, 16 May, 2023 11:30 AM @ MR1 - SSB233 |
| Abstract: | : | This work deals with Network Slicing-based 5G Networks to support the
varying demands of customers and for efficient resource utilization. A
network slice can be defined as a set of network and virtual network
function (VNF) resources deployed across multiple administrative
domains. Here, multi-domain refers to multiple infrastructure providers
spread across different geographic regions. Slice demands and QoS
requirements may vary dynamically, which can be satisfied by scaling the
allotted VNF resources. The VNF scaling problem can be posed as a
time-series forecasting problem that predicts future VNF resources based
on the slice traffic demand. 5G deployments with multiple domains pose
a serious challenge in terms of data privacy since one domain may need
access to the data of another domain for efficient resource allocation
using the conventional forecasting approaches that require data
aggregation.
In this work, we use the federated learning approach in which the training data remains within the respective domains but learns a shared model by aggregating locally-computed updates. We evaluate the applicability of federated settings in VNF scaling using two state-of-the-art deep learning models, Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs). We present a comparison of the performance of the proposed federated system against the centralized system. Additionally, synthetic data in each domain has been generated using Generative Adversarial Networks (GANs) to improve the forecasting results. Next, we wrote a discrete event simulator using a Python-based discrete simulator SimPy to study the performance of our auto-scaling system. Using the simulator, we conducted experiments to compare the performance of scaling and non-scaling systems across various workloads. We also compare our proactive approach's effectiveness against the reactive VNF scaling. |
