Internet of things (IoT) based smart city applications rely on constant data collection and accurate data analytics, yet the fast-changing nature of such data often causes the performance of machine learning models to deteriorate over time. Adaptive learning has been increasingly utilized in these applications in recent years as a viable solution to this problem. Moreover, IoT applications are vulnerable to various security threats due to their large-scale deployment, resource-constrained devices, and diverse protocols. This has led to an increased interest in efficient security and intrusion detection mechanisms tailored for IoT environments. In this chapter, the authors first focus on methods to address the issue of concept drift in time series streaming data for IoT-based smart city applications, such as weather, flood, and energy consumption forecasting, through adaptive learning. Furthermore, the authors examine adaptive learning-based security solutions to various attacks in different domains of the dynamic smart city landscape.
Top1. Introduction And Background: Iot-Based Smart Cities And Adaptive Learning
Recent years have seen an increased interest in learning from continuously evolving stream data, particularly when this data is prone to change and trends frequently (i.e. concept drift). Unfortunately, conventional mining techniques and algorithms are insufficient to solve this problem, as the model's performance degrades even with stationary data, let alone data streams. Continuously adjusting the model manually is ineffective, and as the data size continues to expand, it also becomes impractical. Through constant model iteration and taking advantage of newly available data, the performance of a machine learning model can be improved by rapidly adapting to unforeseen changes in conditions. Adaptive AI models possess the ability to dynamically readjust based on real-time feedback, which is an important aspect of their suitability for unpredictable environments requiring fast response. Adaptive AI solutions are also useful for real-time monitoring and response mechanisms to mitigate the impact of security breaches and intrusion attempts. The purpose of this chapter is to help readers reach a deeper understanding of Adaptive AI, particularly pertaining to its use in Internet of Things (IoT) based smart city applications.
Smart city applications depend on the variety and correctness of large amounts of data collected from IoT devices. Sensing and transmitting data is only the first step, and the bigger challenge is generating actionable intelligence from constantly flowing big data for more efficient resource consumption, planning and preparedness in cities (Kim, Ramos, & Mohammed, 2017). Hence, it is crucial to use stream processing and analytics effectively to fully benefit from wide-scale deployment of IoT devices in smart cities. To this end, distributed stream processing frameworks are important enablers for data analytics (Nasiri, Nasehi, & Goudarzi, 2019).
The relationship between adaptive learning and stream processing for smart city applications has been studied in the literature. For example, a series of experiments to determine whether adaptive learning can improve the performance of stream analytics was conducted by Ku (2018). The experimentation-heavy study utilized six data sets and made extensive use of data. Comparing the speed and accuracy of traditional machine learning classifiers and adaptive ones, the author concluded that adaptive learning provides superior speed and accuracy.
The majority of prior research on adaptive learning on streaming data has focused on classification problems. For instance, researchers (Zliobaite, Bifet, Pfahringer, & Holmes, 2011) developed a model that adapts upon the detection of concept drift. According to the authors, their model is based on uncertainty and dynamic allocation, and responsive to any changes in the distribution of the data set. Others have investigated classifier ensembles and various proposed algorithms to find the best and most efficient classifiers for online classification (Mehta, 2017), (Kuncheva, 2008).
Furthermore, different researchers have approached the problem from various angles. According to one study (Das, Zhong, Stoica, & Shenker, 2014), the batch size of streaming data that is executed affects the throughput and end-to-end latency. Based on this idea, the authors proposed a simple but rigid control algorithm that automatically adjusts the batch size as needed. Another work (Li, Wang, Wang, & Zhou, 2017) addressed the issue of imbalanced streaming data using multi-window-based ensemble classification. Prior to classifying newly arriving streams, their proposed method evaluates the accuracy of the sub-classifiers. When the precision falls below a predetermined threshold, new sub-classifiers are trained. In addition, Zliobaite and Gabrys (2012) addressed the issue of preprocessing in streaming data by considering three scenarios involving adaptive preprocessing, adaptive learning, and adaptive preprocessing and learning in conjunction. They argue that combining adaptive preprocessing and adaptive learning improves the overall performance of machine learning models.