I have developed this model during my internship under University of Galway,Ireland under the supervision of senior researchers of the Institute.
The threat of unwanted actions like invasions and cyber attacks has grown in importance as the reliance on com- puter networks for crucial functions increases. The identification and prevention of such assaults depend heavily on intrusion detection systems (IDS). Due to the complexity and diversity of threats, traditional IDS methods frequently have significant false positive and false negative rates. Hence, to improve the precision and efficacy of IDS, new machine learning approaches are required. In this study, we provide an improved intrusion detection strategy that makes use of RobustScaler, which is a much efficient than other Scaler like MinMaxScaler, it is less affected by outliers than other scaling techniques. By using RobustScaler, machine learning models can be trained on more robust and reliable data, leading to better performance and accuracyoutlier identification with Isolation Forest and Edited Nearest Neighbor (ENN). Isolation Forest is an ensemble-based approach for detecting anomalies that can isolate outliers into their own trees in order to find them in a dataset.By reducing noisy and redundant data points, ENN is a technique used to clean and refine the dataset.Along with RobustScaler that is an efficient tool for normalization. We use a Convolutional Neural Network (CNN) coupled with an LSTM architecture as the classification model following dataset refinement. The CNN-LSTM model can extract both spa- tial and temporal information from network traffic data, allowing it to recognise the complex multiclass attack patterns. The CNN- LSTM model is trained for multiclass intrusion detection using the revised dataset. Using the UNSW-NB15 dataset, we assess our suggested strategy and compare it to other cutting-edge intrusion detection techniques. Experimental findings show that our system detects multiclass incursions accurately while limiting false positives and false negatives. It also achieves much higher accuracy, precision, recall, and F1-score than older methods. Additionally, our method demonstrates robustness in managing a range of assault situations. Further research can be conducted to investigate the scalability, generalizability, and applicability of our approach to other real-world network security datasets
