Résumé: Intrusion detection systems (IDSs) investigate various attacks, identify malicious patterns, and implement effective control strategies. With the recent advances in machine learning and deep learning, designing an efficient IDS has become feasible. Existing Deep learning (DL) methods predominantly suffer from data loss or overfitting while handling class-imbalanced data. Generative adversarial networks (GANs) efficiently solve the overfitting class and overlap problems. However, GANs suffer from “instability” in the training process. Another issue also arises in GANs when the input data does not accurately reflect the actual distribution of the data, such that the generator would produce samples of one or a very small number of classes rather than generating samples for all minor classes at the same time; this behavior is called "mode collapse." To reduce the instability and mode collapse problems and improve the detection accuracy, we developed novel architectures for the generator and discriminator to improve the multi-attack detection problem with a stable training process using ensemble deep learning (EDL) models across various loss functions. To solve the mode collapse problem, we proposed multi-player GANs with teams of multiple generators and discriminators to optimize the learning process while minimizing the chance of mode collapsing. Experimental results over various benchmark datasets show that the developed ensemble-based multi-player GANs (EMP-GANs) outperform the existing methods in terms of quality while maintaining high training stability and minimal chances of mode collapse.

Langue: Anglais