Kernel‐PCA + RBM Feature Extraction with Optimised Naïve Bayes for Intrusion Detection in Impaired Wireless Sensor Networks

Abstract

The high-dimensional communication data generated by wireless sensor networks often contains substantial redundant and irrelevant information, which hampers the effective retention of critical features. Consequently, the characteristics of network impairment states and abnormal intrusion behaviors become intertwined and difficult to distinguish, ultimately compromising the accuracy of intrusion detection. Therefore, this paper studies the method of abnormal intrusion detection of wireless sensor network communication under network impairment. First, global node perception is achieved through the wireless sensor network networking model to obtain high-dimensional communication data. Second, the kernel principal component analysis (KPCA) method is used to perform nonlinear dimensionality reduction on the data, significantly reducing the data dimension and computational complexity while retaining the key information in the data. Subsequently, a restricted Boltzmann machine (RBM) is introduced to extract the deep features of the dimensionality-reduced data to distinguish the feature differences between network impairment states and abnormal intrusions. Finally, a high-precision abnormal intrusion detection is achieved through an optimized naive Bayes classifier. This classifier effectively improves the anti-interference ability under network impairment states by feature weighting and micro conditional probability optimization, highlights key features, and realizes abnormal intrusion detection. The experiment was conducted on a WSN dataset containing 50000 records, simulating a damaged scenario with a 30% packet loss rate and a 40% bandwidth limitation. The results showed that the proposed method reduced the data dimensionality from 90 to 15 dimensions, with a variance retention rate of 94.7%; In the detection of 10 types of attacks, the F1 value reaches 0.92, which is better than CNN (0.60) and association rules (0.62); At a 75% network damage rate, the false positive rate is only 5%, with accuracy and recall rates of 0.94 and 0.86, respectively, and a single sample prediction time of only 0.21 ms. This method maintains high detection accuracy while having low computational overhead and strong robustness, making it suitable for WSN security protection in complex damaged environments.