Unsupervised Learning for Trustworthy IoT
Unsupervised Learning for Trustworthy IoT
复制标题
DOI：
10.1109/fuzz-ieee.2018.8491672
复制
发表时间：
2018-05
期刊：
影响因子：
--
通讯作者：
Nikhil Banerjee;Thanassis Giannetsos;E. Panaousis;C. C. Took-C.
中科院分区：
文献类型：
--
作者：
Nikhil Banerjee;Thanassis Giannetsos;E. Panaousis;C. C. Took-C.
具有各种传感器的物联网(IoT)边缘设备的进步使我们能够利用移动人群感知应用(MCS)来利用各种信息。这种高度动态的环境需要收集来自人类携带的传感器的无处不在的数据痕迹，这带来了新的信息安全挑战；其中之一是数据可信度的保存。在这些设置中需要的是及时分析这些大型数据集，以便对用户报告的正确性产生准确的见解。现有的数据挖掘和其他人工智能方法是从物联网数据中获得隐藏洞察力的最受欢迎的方法，尽管面临许多挑战。在本文中，我们首先在智能和合谋对手存在的情况下对MCS报告的网络可信性进行建模。然后，我们使用真实的物联网数据集，严格评估知名数据挖掘算法在用于物联网安全和隐私方面的有效性和准确性。通过考虑潜在现象的时空变化，我们展示了概念漂移如何伪装攻击者的存在以及它们对聚类和分类过程的准确性的影响。我们最初的一组结果清楚地表明，这些无监督学习算法容易受到对手感染，因此，通过利用先进的机器学习模型和数学优化技术的组合，放大了在该领域进一步研究的必要性。
The advancement of Internet-of-Things (IoT) edge devices with various types of sensors enables us to harness diverse information with Mobile Crowd-Sensing applications (MCS). This highly dynamic setting entails the collection of ubiquitous data traces, originating from sensors carried by people, introducing new information security challenges; one of them being the preservation of data trustworthiness. What is needed in these settings is the timely analysis of these large datasets to produce accurate insights on the correctness of user reports. Existing data mining and other artificial intelligence methods are the most popular to gain hidden insights from IoT data, albeit with many challenges. In this paper, we first model the cyber trustworthiness of MCS reports in the presence of intelligent and colluding adversaries. We then rigorously assess, using real IoT datasets, the effectiveness and accuracy of well-known data mining algorithms when employed towards IoT security and privacy. By taking into account the spatio-temporal changes of the underlying phenomena, we demonstrate how concept drifts can masquerade the existence of attackers and their impact on the accuracy of both the clustering and classification processes. Our initial set of results clearly show that these unsupervised learning algorithms are prone to adversarial infection, thus, magnifying the need for further research in the field by leveraging a mix of advanced machine learning models and mathematical optimization techniques.