CWMAGAN-GP-Based Oversampling Technique for Intrusion Detection

Addressing the challenge of classifiers’ diminished detection performance stemming from class imbalance in industrial control system anomaly detection, this paper introduces a CWMAGAN-GP oversampling model that is based on generative adversarial network. To maintain trainability, this model sets up a non-numerical column output layer using the Gumbel Softmax activation function in the generator and an embedding layer in the discriminator. Meanwhile, it incorporates a multi-head attention mechanism to augment the network's capacity to capture crucial information and enhance feature extraction. Furthermore, the loss function has been augmented with the Wasserstein distance and gradient penalty to improve the stability of its training. In order to evaluate the effectiveness of our suggested model, we performed experiments using the UNSW-NB15 and NSL-KDD datasets. The experimental results conclusively showcase the model's substantial capacity to bolster the detection performance of machine learning detection models.