The growing frequency and complexity of network-based cyberattacks have revealed inherent weaknesses within traditional signature-based intrusion detection systems (IDS), which lack the ability to generalise beyond predefined attack signatures. This opens up an alternative, machine learning (ML) based approach to detection of known and not previously seen threats in a data driven manner. Nevertheless, binary classification tasks dominate among ML-based studies on the UNSW-NB15 benchmark dataset but the more challenging multi-class problem, especially with severe class imbalance, is largely unexplored. This study has two objectives; first, to evaluate and compare the classification performance of five ML algorithms including Random Forest (RF), XGBoost, LightGBM, K-Nearest Neighbours (KNN), and a Multi-layer Perceptron (MLP) on 10-class intrusion detection using the UNSW-NB15 dataset, and second, to investigate the effect of imbalance on percategory detection performance as well as quantify this impact by assessing how SMOTE contributes towards accurate identification of minority attack classes. In this work, we preprocessed the official UNSW-NB15 train/test splits. Mutual information-based feature selection of 20 out of 49 features were retained with label encoding and Min-Max normalisation. This was so that SMOTE was only applied to the training set and do not leak any information from both sides. XGBoost was the most accurate on average with an accuracy score of 98.23% and macro F1-score of 86.14%, at MCC: 0.934 The F1-score of rarest class Worms increased by 31.11 percentage points from 22.22% to 53.33% with SMOTE. Results solidify that tree-based ensemble techniques, especially XGBoost, are appropriate for multi-class intrusion detection on UNSW-NB15. Future work must explore federated learning frameworks, real-time streaming evaluation, and transfer learning across complementary benchmark datasets