AI applications in networking

- [Narrator] AI and machine learning, ML, has gained much attention in many ID products and services, but what are the applications when applying to data networks? A data network in a form of an enterprise network, for example, needs to interconnect network devices such as client devices, routers, and switches, and support various network services, which are the applications run on servers at application layer for connecting or providing information to users in different locations. However, there is much more work behind the scenes for administration, operations, and management. A network also needs to safeguard and manage the devices and services. All these require automation and autonomy, which can be provided through AI and ML, but we'll refer to AI and ML as AI for simplicity. For example, at the edge router or core switch of the enterprise network, AI can be deployed to learn the pattern of the data traffic so that we can detect if they are anonymous events resulting from misconfigurations, policy violations, malicious attacks, congestion, et cetera. After the firewall, we can mirror the traffic going through the enterprise network to the network intrusion detection system, IDS, to detect and prevent malicious attacks by inspecting the traffic data and using a deployed AI model to classify this traffic to see which and when data traffic is malicious. At the distribution layer of an enterprise network, AI can help optimize the traffic flows by offloading some network devices. The policy violations may be detected at this distribution layer and at the access layer of an enterprise network. It would be good to know the health status of all devices and the network, and if something is wrong, we want to know the root causes of the problem. At the wireless local area network controller, a management portal can visualize the network health information by analyzing the telemetry and the management data sent by wireless access points and the wireless client devices. This application applies to the wired ethernet networks, as well. With the data, AI can help predict when a failure may occur. With the application traffic information collected over time, we can use AI to predict the application usage in the future. For the network services or applications provided by a server, there are many application-specific monitoring and optimization tasks that can be assisted by AI techniques. For example, dynamic load balancing can be achieved by reinforcement learning, where user tasks requesting for a server resource can be optimally assigned to be handled by the best server, considering the current workload and computing resources of all servers. An enterprise wide-area network, enterprise WAN, has geographically distributed offices connected through routers. Among those routers, AI-enabled adaptive routing can be used to make routers established and optimal routes in response to the changing user demands, bandwidth, delays, and other quality of service requirements. This network architecture may also be applicable to the data center networks. While we are focusing on our discussion on networking, we should realize that AI has been broadly considered in communication systems. For example, AI can enable the optimal traffic allocation while there are multi-link communications, where devices can transmit and receive data over multiple radio interface concurrently. A future Wi-Fi router may determine the best percentage of the traffic flow to be used for each interface. In a cellular network such as 5G, we can use AI for beam management for better steering of directional beings with at antenna arrays. These applications can be transferrable to various network paradigms, where many real world network setups follow. For example, on a vehicular network, the network protocol needs to adapt to dynamic topology changes. While time sensitive and guaranteed message delivery can be achieved between vehicles to announce accidents, traffic congestion, obstacles, road distortion, and conditions. On an optical network, we would need to minimize deflection actions and selected the best alternate path while minimizing packet drops. On a delay or disruption-tolerant network, we can use AI techniques to maximize the number of packets delivered to a destination. Software-defined networking, or SDN, is an important network management approach that has been adopted by many networks on an SDN-based network. SDN controllers will need to determine the best routes and send routing tables to switches, which follow the received tables to forward data packets. On an internet of things or a sensor network, we may want to maximize network lifetime based on the constrained compute and energy resources on the sensing devices. On a space-air-ground integrated network, where satellite networks, aero platforms are to be integrated with retrial networks, we may want to monitor and predict the network performance and allocate resources optimally with the possibility of planning the trajectory of aero platforms, such as drones, on a 5G network to deal with the complexity of network functions. The 5G standardization body, 3GPP, has introduced the network data analytics function, NWDAF, which collects and analyzes data from network functions and the network devices, and uses AI algorithms to facilitate control automation, self-healing, and performance optimization at the 5G core. As we can see, AI can be applied to many networking scenarios. We'll continue our discussion on some representative tasks in network management, optimization, security, and resilience, which could be applicable to any network.