Imputed-VAE IIoT-2021

Class imbalance and anomalous data points are critical challenges in many real-world classification tasks, often leading to biased or inaccurate predictive models. This paper presents an integrated 3 phase data transformation framework that combines Isolation Forest based outlier detection, neural network imputation, and Variational Autoencoder driven class balancing to address these issues effectively for IIoT intrusion detection. Initially, Isolation Forest identifies feature-level outliers separately within each class. Subsequently, neural networks are trained on clean data to impute outlier values, preserving class specific characteristics. To mitigate severe class imbalance, VAEs are employed for minority class oversampling and majority class undersampling based on reconstruction and KL divergence errors. The downsampled, class-balanced dataset demonstrates matching classification performance with the WUSTL IIoT 2021 dataset across multiple algorithms, achieving near-perfect accuracy, precision, recall, and F1-score. The proposed pipeline has three key benefits: it enhances model robustness, offers a scalable solution for imbalanced and noisy datasets common in industrial IoT security applications, and functions as a representative, small-scale version of the original dataset.