Department of Computer Science and Information Technology, Central University of Jammu, Rahya Suchani, Jammu and Kashmir, 181143, India.
Leaders Institute, Woolloongabba, Brisbane, QLD, -4102, Australia.
Sci Rep. 2022 Jul 18;12(1):12247. doi: 10.1038/s41598-022-16261-9.
The next whooping revolution after the Internet is its scion, the Internet of Things (IoT), which has facilitated every entity the power to connect to the web. However, this magnifying depth of the digital pool oil the wheels for the attackers to penetrate. Thus, these threats and attacks have become a prime concern among researchers. With promising features, Machine Learning (ML) has been the solution throughout to detect these threats. But, the general ML-based solutions have been declining with the practical implementation to detect unknown threats due to changes in domains, different distributions, long training time, and lack of labelled data. To tackle the aforementioned issues, Transfer Learning (TL) has emerged as a viable solution. Motivated by the facts, this article aims to leverage TL-based strategies to get better the learning classifiers to detect known and unknown threats targeting IoT systems. TL transfers the knowledge attained while learning a task to expedite the learning of new similar tasks/problems. This article proposes a learning-based threat model for attack detection in the Smart Home environment (SALT). It uses the knowledge of known threats in the source domain (labelled data) to detect the unknown threats in the target domain (unlabelled data). The proposed scheme addresses the workable differences in feature space distribution or the ratio of attack instances to a normal one, or both. The proposed threat model would show the implying competence of ML with the TL scheme to improve the robustness of learning classifiers besides the threat variants to detect known and unknown threats. The performance analysis shows that traditional schemes underperform for unknown threat variants with accuracy dropping to 39% and recall to 56.
继互联网之后，下一个重大变革是其分支物联网 (IoT)，它使每个实体都具有连接互联网的能力。然而，这极大地加深了数字领域的深度，为攻击者提供了可乘之机。因此，这些威胁和攻击已经成为研究人员关注的首要问题。机器学习 (ML) 具有广阔的前景，它是检测这些威胁的主要手段。但是，由于域的变化、不同的分布、长时间的训练以及缺乏标记数据，基于通用 ML 的解决方案在实际实施中对于检测未知威胁的效果逐渐下降。为了解决上述问题，迁移学习 (TL) 已经成为一种可行的解决方案。受此启发，本文旨在利用基于 TL 的策略来改进学习分类器，以检测针对物联网系统的已知和未知威胁。TL 将在学习任务中获得的知识转移到加速新的类似任务/问题的学习中。本文提出了一种基于学习的智能家居环境 (SALT) 攻击检测威胁模型。它利用源域（标记数据）中已知威胁的知识来检测目标域（未标记数据）中的未知威胁。所提出的方案解决了特征空间分布或攻击实例与正常实例的比例或两者之间的差异。所提出的威胁模型将展示 ML 与 TL 方案相结合的可行性，以提高学习分类器的鲁棒性，同时还可以检测已知和未知的威胁变体。性能分析表明，传统方案对于未知威胁变体的性能较差，准确率下降到 39%，召回率下降到 56%。
