Machine Learning Aided Network Management: Awareness, Connectivity, and Security

Abstract

As the generation of network technology advances, a new paradigm for network management has been required to satisfy the demanding network performance for next-generation networks. Recently, with the advance of machine learning (ML) techniques, ML is exploited in various fields and such works show impressive performance compared to conventional methods. In the same context, ML has been introduced in network management to manage the complex systems through various approaches. In this dissertation, we propose deep learning and reinforcement learning based network management techniques for improving environment awareness, connectivity, and security for next-generation networks.

In the first study of this dissertation, challenging radio map estimation problems for multiantenna channel state information (CSI) are tackled using deep learning. The desired CSI is predicted for arbitrary locations in a geographical area based on CSI measurements collected at a small number of sampling locations. Such maps can be used to significantly reduce the overhead associated with CSI acquisition. A novel deep architecture is proposed, consisting of an encoder/decoder pair for transforming high-dimensional CSI features to lower-dimensional embeddings, and a deep embedding interpolator for exploiting the spatial dependency of the multiantenna CSI.

In the second study of this dissertation, we propose a deep reinforcement learning (DRL)-based routing optimization on an software-defined networking (SDN). In the proposed method, the DRL agent learns the interdependency between the traffic load of network switches and the network performance, and decides an optimal set of link weights to make a balance between the end-to-end delay and packet losses of the network. The SDN controller determines the routing paths using the set of link weights and installs the flow-rules on the SDN-enabled switches. To overcome an extensively long learning process of DRL in a case of topology change, we develop an M/M/1/K queue-based network model and perform the learning process of DRL using the network model in an offline manner until it is converged. The simulation results demonstrate the proposed routing method outperforms a conventional hop-count routing and a traffic demand-based RL algorithm in several network topologies.

In the third study of this dissertation, we consider a jamming attack to an unknown wireless network where no \textit{a priori} network information is provided to the jammer except the radio frequency signal information acquired by overhearing the shared wireless channel. To increase the impact of a jamming attack on an unknown network, we propose a reinforcement learning based beamforming attack strategy. In the proposed attack, a jammer learns the beam direction and angle width to maximize the impact of jamming using the multi-armed bandit technique. As a reward for MAB, we develop a new metric that can quantitatively evaluate the impact of a jamming attack by measuring the statistical change of the channel busy times before and after each attack. Through extensive simulations, we evaluate the performance of the proposed jamming strategy in an unknown wireless network.

Through the studies in this dissertation, the proposed network management methods improves service performance in wired or wireless networks. We expect that the proposed network management greatly contribute to further improve service performances of the next-generation networks.