Machine Learning in Networking

Parham Alvani

@1995parham

Introduction

• Networks needs more intelligence to deploy
• Traditional networks are distributed
• Each node, such as router or switch, can only view and act over a small partion of the system.
• Learning from the nodes that have only a small partial view of the complete system to perform contol beyond the local domain is very complex

Introduction

• SDN decouples the control plane and the data plane
• The network resources in SDN are managed by a logically centralized controller

Software Defined Networking

Machine Learning in SDN

• Traffic Classification
• Routing Optimization
• QoS/QoE Prediction
• Resource Management
• Security

• Traffic classificiation is an important network function, which provides a way to perform fine-grained network management by identifying different flow types.
• In general supervised and semi-supervised learning algorithms can be used
• Deep Packet Inspection (DPI) is a common method to label traffic flows, but it incurs high computational cost when a large number of traffic flows are labeled

• Routing is a fundamental network function
• Supervised Learning for obtaining the optimal huristic-like routing solution
• RL algorithms without labeled taining datasets can have flexible optimization targets
• The state space is composed of network and traffic states

• QoS prediction aims to discover the quantitative correlations between KPIs and QoS parameters
• QoS parameters are generally continuous data (Regression)
• As an example here we can use NN-Model for network delay estimation instead of M/M/1 model
• QoE prediction aims to discover the quantitative correlations between QoS parameters and QoE values
• QoE values are generally discrete data (Classification)

• Efficient network resource management is the primary requirement of network operators to improve network performance
• The data plane resource allocation problem is generally considered as a decision-making task.
• RL and ML-based game theory are two effective approaches
• RL in single-tenancy
• ML-based game theory in multi-tenancy
• The mapping between the resource consumption and control message rate is very important for control plane resource allocation

• An Instrusion Detection System (IDS) is a device or software application and its objective is to monitor the events in a network system and identify possible attacks
• signature-based IDS and anomaly-based IDS
• Machine learning methods are widely used in anomaly-based IDS by trainning a model to identify normal activities and intrusions
• Intrusion detection problem can be considered as a classification task

• High-quality Training Datasets
• Distributed Multi-controller Platform
• Improving Network Security
• Cross-layer Network Optimization
• Incrementally Deployed SDN