Abstract
Cyber-attacks are becoming common in linked environments these days, which means that any devices, no matter how big or tiny, are vulnerable to them. In order to proactively anticipate and mitigate network threats, it becomes imperative to design Intrusion Detection Systems (IDS) for these interconnected environments. As such, a number of cutting-edge projects concentrate on developing IDS models by applying Deep Learning (DL) techniques. This change is a result of the shortcomings of traditional models, which primarily provide static IDS systems and underline the need for more sophisticated and dynamic IDS systems. With this goal in mind, we create the Parallel Attention Self Supervised based Convolution Neural Network (PASS-CNN), a brand-new self-supervised dynamic IDS model. Three sequential processes make up the designed self-supervised IDS model: feature extraction and dynamic aggregation, data pre-processing, and dynamic IDS. First, the traffic that is obtained from the network is pre-processed in terms of data normalisation, data smoothing, and data encoding, in that order. After the traffic has been pre-processed, it is made available for feature extraction using the CIC-flow metre tool. To decrease complexity and scalability, the collected features are then dynamically aggregated by creating dynamic aggregation rules using the Multi Agent Deep Reinforcement Learning (MADRL) algorithm. In order to effectively detect intrusions, the characteristics are finally controlled to the suggested PASS-CNN model, which is made up of encoder, decoder, and parallel attention modules. Using benchmark datasets like the UNSW-NB15 and CICIDS-2017 datasets, respectively, the efficacy of the proposed model is compared to several other IDS models in terms of assessment metrics including accuracy, precision, recall, F1-score, and false positive rates. The evaluation's findings demonstrate that the suggested results outperform the current ones.