Multi-Label Classification for Drift Detection in IoT Data Streams

Abstract

With the prevalence of Internet of Things (IoT) systems, data is exponentially growing, resulting in data streams. Data streams are massive, potentially non-stop, and arrive at high speed. These systems process IoT data streams in a non-stationary manner, making them susceptible to concept drift occurrence and class imbalance. Concept drift occurs as a result of the change in the underlying distribution over time, while class imbalance occurs when some class distribution is uneven. This paper proposes an Implicit Drift Detection model with Multi-Label kNN (IDD-MLkNN), aimed at addressing concept drift in multi-label classification for IoT data streams. While the model is applicable across various domains, its performance was specifically assessed using two IoT datasets — Bot_IoT and Edge_IIoTset — associated with intrusion detection systems. Applying the proposed model to oversee IoT network traffic offers practical advantages, potentially reducing the time and expenses of re-examining attack data. The evaluation was conducted for sudden and gradual concept drift scenarios. Experimental results show the superiority of the IDD-MLkNN over other well-known multilabel classification models in different performance measures such as the Subset Accuracy, Accuracy, Hamming Score, and F-measure. However, it was less efficient in terms of Evaluation Time compared to other multi-label classification methods.