Feature Engineering using Bio-Inspired Methods for the Internet Data Analytics

互联网流量的准确分类一直是互联网服务提供商长期关注的问题，主要有两个原因：网络安全和网络资源分配。随着互联网流量的增长，在应用的数量和类型方面，传统的互联网数据分析识别方法面临着严峻的挑战，因为更新的和/或恶意的应用使用随机选择的端口进行通信，使得它们越来越难以识别。因此，利用机器学习、进化方法和数据挖掘技术对互联网流量的统计行为进行分析得到了学术界和工业界的广泛关注。为此，必须对互联网数据进行捕获和分析。从通信网络提取的特征非常多样化，数据集通常是大的和多维的。因此，有效地转换和计算重要特征以进行充分的数据分析是需要重点关注的重要研究任务。 这项研究计划的目标是探索创新的算法，利用生物启发和机器学习方法，包括人工免疫系统和遗传算法，对互联网数据进行特征工程和分析。我们将设计和分析一个由特征工程模块、分类器模块和再培训机制组成的体系结构。我们将探索围绕生物启发算法建立的新方法，包括基于类的和独立于类的遗传映射，以及基于人工免疫系统的异常检测，同时考虑到训练数据集中的不平衡数据和缺乏足够的标记。我们将探讨所提出的算法在计算机网络中的两个重要应用中的能力：(A)用于网络安全的入侵检测系统和异常检测；以及(B)用于资源分配和服务质量保证的互联网流量的协议识别。

Accurate classification of the Internet traffic has been a permanent concern for the Internet Service Providers for the two main reasons: network security; and network resource allocation. With the growth of the Internet traffic, in terms of the number and type of applications, traditional identification methods for the Internet data analytics face serious challenges when the newer and/or malicious applications use randomly selected ports to communicate making them increasingly difficult to identify. Thus, analysis of the Internet traffic based on its statistical behaviors using machine learning, evolutionary methods, and data mining techniques have gained much attention in academia and industry. For this, the Internet data must be captured and analyzed. The features extracted from the communications networks are very diverse and the datasets are typically large and multidimensional. Consequently, efficient transformation and computation of the significant features for adequate data analytics are important research tasks to focus on. The objective of this research program is to explore innovative algorithms for feature engineering and analysis of the Internet data using bio-inspired and machine learning approaches including Artificial Immune Systems and Genetic Algorithms. We will devise and analyze an architecture comprising of a feature engineering module, a classifier module, and a re-training mechanism. We will explore new methods build around bio-inspired algorithms including class-dependent and class-independent genetic-based mapping, as well as artificial immune systems-based anomaly detection while considering imbalanced data and lack of sufficient labeling in the training dataset. We will explore the competency of the proposed algorithms in two important applications in computer networking: (a) intrusion detection systems and anomaly detection for network security; and (b) protocol identification of the Internet traffic for resource allocation and quality of service assurance.