Deep generative models for network data synthesis and monitoring

Abstract

Measurement and monitoring are fundamental tasks in all networks, enabling the down-stream management and optimization of the network. Although networks inherently have abundant amounts of monitoring data, its access and effective measurement is another story. The challenges exist in many aspects. First, the inaccessibility of network monitoring data for external users, and it is hard to provide a high-fidelity dataset without leaking commercial sensitive information. Second, it could be very expensive to carry out effective data collection to cover a large-scale network system, considering the size of network growing, i.e., cell number of radio network and the number of flows in the Internet Service Provider (ISP) network. Third, it is difficult to ensure fidelity and efficiency simultaneously in network monitoring, as the available resources in the network element that can be applied to support the measurement function are too limited to implement sophisticated mechanisms. Finally, understanding and explaining the behavior of the network becomes challenging due to its size and complex structure. Various emerging optimization-based solutions (e.g., compressive sensing) or data-driven solutions (e.g. deep learning) have been proposed for the aforementioned challenges. However, the fidelity and efficiency of existing methods cannot yet meet the current network requirements.

The contributions made in this thesis significantly advance the state of the art in the domain of network measurement and monitoring techniques. Overall, we leverage cutting-edge machine learning technology, deep generative modeling, throughout the entire thesis. First, we design and realize APPSHOT , an efficient city-scale network traffic sharing with a conditional generative model, which only requires open-source contextual data during inference (e.g., land use information and population distribution). Second, we develop an efficient drive testing system — GENDT, based on generative model, which combines graph neural networks, conditional generation, and quantified model uncertainty to enhance the efficiency of mobile drive testing. Third, we design and implement DISTILGAN, a high-fidelity, efficient, versatile, and real-time network telemetry system with latent GANs and spectral-temporal networks. Finally, we propose SPOTLIGHT , an accurate, explainable, and efficient anomaly detection system of the Open RAN (Radio Access Network) system. The lessons learned through this research are summarized, and interesting topics are discussed for future work in this domain. All proposed solutions have been evaluated with real-world datasets and applied to support different applications in real systems.