PBPCSANet: A poisson bias loss-based parallel convolutional stacked autoencoder network for minority class attack detection

Minority class attack detection is used to identify rare attacks, like malicious activities or intrusions in highly imbalanced databases, and prevailing schemes of attack detection mostly fail to detect minority attacks because of the limited representation of rare instances. In this context, Deep Learning (DL) schemes employed in identifying minority class instances achieved a high accuracy rate. However, these approaches tend to overfit and struggle in learning the patterns of minority attacks precisely. Therefore, a novel Poisson Bias Loss-based Parallel Convolutional Stacked Autoencoder Network (PBPCSANet) is introduced to detect minority class attacks. First, the input data is normalized by applying Sigmoid normalization. Further, the features from the normalized data are selected, and then the resultant is augmented using Adaptive Synthetic Sampling (ADASYN). Afterwards, the minority class attack is detected by exploiting the PBPCSANet, which is advanced by adjusting the learning rule of the Parallel Convolutional Stacked Autoencoder (PConvSA-Net) with Taylor Poisson Bias Loss (TPoBL). Here, the TPoBL is presented by the amalgamation of Poisson loss, Taylor series, and Mean Bias Error (MBE). Moreover, the PBPCSANet model achieved a minimal False Positive Rate (FPR) of 3.631%, higher True Negative Rate (TNR), Accuracy, True Positive Rate (TPR), precision, and F1-score of 96.369%, 96.690%, 97.228%, 95.998%, and 96.609%, respectively.