Efficient Intrusion Detection Using Multi-Player Generative Adversarial Networks (GANs): An Ensemble-Based Deep Learning Architecture

Intrusion Detection Systems (IDSs) investigate various attacks, identify malicious patterns, and organize and implement effective control strategies. With the recent advances in machine learning and deep learning, designing an efficient IDS has become feasible. It is usually the case that training data has a much higher number of normal data samples than intrusion data points in real-world intrusion datasets; this is called a class imbalanced distribution problem. However, when learning from imbalanced data, the performance of deep learning-based methods drops considerably despite their high accuracy when learning from large amounts of data. Many studies have examined imbalanced data; however, existing methods predominantly suffer from data loss or overfitting. Generative Adversarial Networks (GANs) can solve the overfitting and class overlap problems by generating synthetic data like the existing ones. However, GANs suffer from the training instability problem, leading to high training oscillations, which results in inaccurate predictions. It is crucial to ensure that generators and discriminators perform equally well. When either of these individuals outperforms the other, the entire training process becomes unstable, and no useful information is acquired. And during training, different samples will be generated each time. Another issue also arises when the input data does not accurately reflect the actual distribution of the data. By doing so, the generator would only be able to produce samples of one or a very small number of classes rather than simultaneously make samples for all minor classes. In this case, a mode collapse or drop occurs. To reduce the problem of instability and mode collapse, we developed different architectures for Generator and Discriminator. We developed a new architecture for generative and discriminative learning to improve multi-attack detection with a stable training process using ensemble convolutional neural networks (CNNs). For our developed model, we have also used various loss functions. Experiments have shown that focal loss can enhance minority class detection. When using the mean squared loss function, the detection rate improved for the discriminator, while the deep feature representation improved through the binary cross-entropy loss function. Additionally, we developed a Multi-player GAN with ensemble architecture to reduce the mode collapse problem and have a stable intrusion detection model.