ASTriA: Adaptive Semi-Supervised Tri-Learning Approach for Scalable Malicious URL Detection

Maintaining a large, current collection of labeled malicious URLs is expensive. Blacklists struggle with newly generated addresses, while purely unsupervised detection can flag harmless URLs without reliable supervision.

ASTriA investigates adaptive semi-supervised tri-learning: multiple learners help assign reliable pseudo-labels to unlabeled samples, expanding a smaller labeled training set over successive iterations.

A limited labeled collection anchors learning, while a larger unlabeled pool supplies additional examples. Lightweight lexical and host-based features keep the representation practical for URL classification.

The published abstract describes iterative high-confidence pseudo-labeling with interpretable classifiers including Random Forest and k-Nearest Neighbors. The tri-learning design uses collaboration among learners to reduce dependence on manually supplied labels.

The IEEE publication record confirms the title, five authors, TENSYMP venue, and pages 1–5. The public abstract supports the labeled/unlabeled workflow, pseudo-labeling approach, feature families, and named classifiers.

Numerical accuracy, dataset sizes, learner configurations, and confidence thresholds are not reproduced here because those details were not available in the accessible abstract.

When applying this method, inspect how pseudo-label errors propagate, how confidence is calibrated, and how performance changes as URL distributions evolve. These are practical considerations for interpreting semi-supervised systems, not additional experimental claims about ASTriA.