A Mirror Environment to Produce Artificial Intelligence Training Data

Di Li,Haruhisa Nozue,Kenichi Tayama,Kazuak Akashi

doi:10.1109/access.2022.3154825

Di Li, Haruhisa Nozue + Show 2 more

Open Access

https://doi.org/10.1109/access.2022.3154825

Copy DOI

Abstract

With the increasing maturity of artificial intelligence (AI) technology, business automation technology has also become a trend. Particularly, network operation and maintenance (O<monospace>&</monospace>M) is expected to soon become automated and more efficient. However, the automation of O<monospace>&</monospace>M is hindered by the lack of network failure data and the cost of collecting data. We thus propose an approach to build a low-cost environment that can produce the same data as the actual production environment and use tools such as chaos engineering to generate training models for fault data. This paper attempts to build the underlying physical network layer using a low-cost single-board computer Raspberry Pi instead of an expensive PC server, while keeping the virtual network layer the same and performing fault simulation, data collection, and AI model training on the constructed virtual network layer. A comparison of the accuracy of the trained AI models verifies the feasibility of replacing the traditional PC server with an inexpensive Raspberry Pi device while keeping the structure and services of the virtual network layer unchanged. Also, a brief comparison with existing techniques is discussed. Our proposed approach solves the problem of insufficient data for AI training while reducing cost and risk.

Highlights

W ITH the rapid development in artificial intelligence (AI) technology, automation is being used in more areas [1] such as robotic process automation and cyber defense [2]
A method is needed to accurately determine whether the AI trained in the mirror environment is the same as the AI trained in the target production environment
Collecting data is challenging in the real target production environment, so the mirror environment is needed to collect data for AI training

Summary

Introduction

W ITH the rapid development in artificial intelligence (AI) technology, automation is being used in more areas [1] such as robotic process automation and cyber defense [2]. 1) Data augmentation Mikołajczyk and Grochowski proposed a method to increase the original image data by using a series of transformations such as orientation change and color transformation to obtain more training data [7]. This simple and efficient method of increasing training data is widely used in image recognition This method is often used in image processing, care must be taken to preserve the dataset’s features while adding data. Network fault data is often difficult to obtain and use to identify learning features [8]. It does not fundamentally solve the problem of insufficient or nonexistent data

Objectives

Methods

Results

Conclusion