Rule-Based Outlier Detection with a Modified Variational AutoEncoder for Enhancing Data Accuracy in Wireless Sensor Networks

In wireless sensor networks (WSNs), a number of outlier detection (OD) methods have been established over time to identify data that do not match the rest of the data. These data are anomalies, outliers, or irregularities because they result from hostile attacks or sensor node malfunctions. OD is difficult due to the classification process and the challenges in feature development. The limitations of conventional OD include insufficient feature learning, noise, excessive dimensionality, and a lack of training data. Unsupervised outlier detection (UOD) in WSNs is a new research area that has been ascertained to perform exceptionally well in terms of detection accuracy. The fact that prior knowledge of abnormal data is unknown, i.e., the ground truth of data is unknown for the classification of outlier instances, is a substantial barrier in UOD. This made it possible for researchers to examine flexible deep learning methods for domain-specific UOD, like Variational AutoEncoder (VAE). In order to achieve an appropriate categorization of anomalous data, this research suggests using a Modified Variational AutoEncoder (ModVAE) model with a rule-based approach. In the first stage of the proposed system, an adaptive VAE model is developed to produce modified latent space vector samples suitable for reproducing the data with reasonable reconstruction loss, and in the second stage, fuzzy rules are generated to precisely classify abnormal data with fewer false positives. The suggested model's efficiency is demonstrated with experiments done using benchmark datasets. The experimental results show improved performance when compared with other approaches available in the literature.