Robust Deep Learning against Corrupted Data in Cognitive Autonomous Networks

Kajo, Marton; Schnellbach, Janik; Mwanje, Stephen S.; Carle, Georg

doi:10.1109/noms54207.2022.9789774

Cited by 1 publication

(1 citation statement)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…CFs are usually trained on complete and accurate datasets, so they may not be robust against such data corruption. In [105] researchers showed that a single corrupted input to an ML model, which is part of an ML model chain, can generate multiple corrupted outputs with each step in the chain multiplying the effect. In networks, multiple CFs often share the same input control parameter (ICP), implying that missing or corrupted data for the ICP will affects the learning of all the CFs.…”

Section: Appendicesmentioning

confidence: 99%

Coordination, Trust and Orchestration of Cognitive, Open, Multi-Vendor Network Automation Functions

Banerjee,

Carle

2023

2023 IEEE Conference on Standards for Communications and Networking (CSCN)

View full text Add to dashboard Cite

Over the past few years, network traffic has grown exponentially with the rise in mobile communication network users and services. This, in turn, increased the operational complexity and hiked the capital and operational expenses of mobile network operations. Present mobile networks are highly complex systems comprising automation mechanisms that leverage numerous technologies. These technologies and automation tools have been developed to deal with the configuration, optimization, and troubleshooting of mobile network operations, facilitating an efficient network management. Self-Organizing Networks (SONs) are an example of one such technology that aims to decrease human intervention by automating network management tasks. Although SON is used widely as an automation tool, it has some shortcomings; particularly, maintenance and upgrade of a rule-based management system like SON are challenging. To address these shortcomings, research is aimed at incorporating machine learning (ML) capabilities in SON and introduced a new management paradigm called cognitive autonomous networks (CANs). Like SON, CAN consists of multiple network automation functions called cognitive functions (CFs). A CF is an intelligent network function that is responsible for managing specific optimization related tasks. Throughout this thesis, we use the following CFs that are relevant for the use cases covered in this thesis: mobility load balancing (MLB), mobility robustness optimization (MRO), coverage and capacity optimization (CCO), and energy savings (ES). These functions have already been standardized by both the network operator community as well as the standardization bodies for SON use cases.These CFs operate on the same radio network in parallel and learn to optimize the same set of radio network parameters. Since they work independently but simultaneously try to adjust the same set of parameters, conflicts are likely to happen among them. These conflicts need to be resolved to keep the network stable and the network operations uninterrupted. The conflicts can be resolved by introducing some coordination among the CFs. In this thesis, we evaluate different paradigms of coordination for CAN and establish that centralized coordination is the most beneficial for CAN. In the first part of the thesis, we propose solutions for centralized coordination for conflict resolution among the CFs. These solutions are based on the Nash Social Welfare Function (NSWF) and a more advanced version of NSWF, the Fisher Market Model, combined with Eisenberg-Gale optimization. Our proposed solutions are designed to serve the combined interest of all the CFs while resolving a conflict. However, when compared between these two approaches, we find that the second one provides a significant improvement over the first one regarding overall network performance.Keeping with the current trend of making the network management system more open, flexible, and agile, in this thesis, we consider an open, multi-vendor CAN where different Finally, I sincerely thank Saya...

show abstract

Section: Appendicesmentioning

confidence: 99%