Internet of Things (IoT) intrusion detection faces challenges due to complex, high-dimensional, and redundant data, making it difficult to capture deep features and respond quickly. To address this, a deep learning-based method is proposed. IoT data is segmented using a sliding window, and principal component analysis (PCA) reduces dimensionality by extracting key features. The processed data is then fed into a deep belief network (DBN), which utilises stacked restricted Boltzmann machines (RBMs) to learn high-level representations for identifying complex intrusions like DDoS, phishing, and DoS attacks. To enhance performance, a bacterial colony optimisation algorithm adaptively optimises the DBN's hidden layer weights and biases. Experiments show that a DBN with five hidden layers and eight units per layer achieves optimal detection, with delays under 5 ms, intrusion identification accuracy over 98%, and system stability above 95%. While performance may fluctuate with extremely high-dimensional dynamic streams and applicability in other IoT domains requires further validation, this method demonstrates strong potential for improving IoT security and reliability.
Wang et al. (Thu,) studied this question.