Deep Learning Methodology for Detecting Breaches to Improve Security in Smart Cities

Modern societies embrace the digitization of nearly every facet of life as a promising, long-term advancement. However, the digital realm, much like physical spaces, is fraught with its share of vulnerabilities and perils. Notably, the rapid and sophisticated evolution of smart cities has brought about substantial transformations in daily existence. Within this context, deep learning (DL) has exhibited remarkable efficacy in the realm of smart technology powered cities. This paper presents an effective intrusion detection (ID) system employing DL, with a multistage approach. Three well-established data sources—GPRS, CIDDS001, and UNSW-NB15—each encompassing subsets of potential attack classes that could disrupt IoT-powered smart cities, serve as foundational data. After collection, these data undergo a pre-processing step that includes normalization, followed by feature extraction using an autoencoder. This process encompasses all types of features, both numerical and nonnumerical. Subsequently, feature selection is executed through random forest (RF), aiming to reduce feature dimensionality and retain only those crucial for enhanced performance. The final stage involves utilizing a stack of Restricted Boltzmann Machines (RBMs) for class prediction. Results from our experiments demonstrate that the proposed model (RF-RBM) outperforms several state-of-the-art models across multiple metrics, boasting recall (0.96), accuracy (0.95), specificity (0.97), and detection rate (0.95).