Intrusion-detection replication
A from-scratch replication of two papers on the UNSW-NB15 network-intrusion dataset: ten classifiers across seven feature scalings.
Sole author.
- Best score: MLP with a power transform: 0.989 (source: UNSW-NB15 replication results, 2025)
- Classifier × scaling combinations: 70 (source: UNSW-NB15 replication results, 2025)
- Papers replicated: 2 (source: UNSW-NB15 replication results, 2025)
- Classifiers compared: 10 (source: UNSW-NB15 replication results, 2025)
Select the features
Elastic-net regression chooses which of the flow features carry signal. The regularisation strength was picked on the error curve (optimal alpha 0.021).
See the classes
Interactive 3-D PCA, t-SNE and K-means clustering show how far the attack types overlap with normal traffic and with each other.
Measure the distances
Mahalanobis distances between class centroids, recomputed under each of the seven scalings, show how a scaler stretches or collapses the space between classes.
Benchmark everything against everything
Decision tree, Naive Bayes, logistic regression, k-nearest neighbours, SGD, SVM, MLP, XGBoost, bagging and a voting ensemble, each under no scaling, min-max, robust, standard, quantile, power and normaliser.
Profile the features
From the second paper: probability distributions, skewness, kurtosis, Pearson correlation and gain ratio per feature, then a complexity comparison of decision tree, Naive Bayes, logistic regression, a neural network and expectation maximisation.