A Dual-Model Anomaly Detection Algorithm for non-linear stream data in Smart City Environments
Anthony J. Bustamante, Sarah Asad, Daniela Nicklas, Brent Lagesse · 2024
In this paper, we introduce a complementary and straightforward mechanism for anomaly detection tailored for smart city infrastructures, utilizing a combination of regression algorithms. Our methodology employs two distinct regression models to generate future predictions from a given dataset. The primary model is crafted to yield high-fidelity predictions, while the secondary model is purposefully designed to introduce a degree of noise. Both models work together as a defense against Flooding attacks through the detection of abnormal levels of data inflow (detection of outliers). We calculate the alignment cost, or Euclidean distance, between the predictions from these two models, establishing a threshold against which real future traffic can be evaluated. The alignment cost or euclidean distance of the actual traffic is computed in relation to the high-quality predictions and then compared with the established threshold to pinpoint anomalies. Through experimentation with various regression algorithms, including linear regression, support vector regression, decision trees, etc., we identified an optimal combination for peak performance. Our assessments, grounded in comprehensive smart city datasets, center on the process of transforming complex non-linear data into a more appropriate form to detect anomalous data points. Conclusively, the dual-model anomaly detection framework we propose stands out as an invaluable tool in defending smart city infrastructures from data irregularities and potential threats, highlighting the criticality of bespoke solutions in contemporary urban digital environments.