Drift-Aware Retraining Matches Periodic Retraining Accuracy for Malware Classifiers at a Fraction of the Compute

arXiv 2608.13465·low signalXBlueskyLinkedInCopy link

This study compares three concept-drift detectors — a novel One-Class SVM approach, Minibatch K-Means, and Maximum Mean Discrepancy — across four malware classifiers (MLP, Random Forest, SVM, XGBoost) under static, periodic-retraining and drift-aware-retraining regimes, with Pareto front analysis of the accuracy/efficiency tradeoff. All three detectors reach classification accuracy comparable to blanket periodic retraining while requiring substantially fewer models to be retrained, and the OCSVM technique generally outperforms MK-Means and MMD. The practical read for anyone running scheduled retraining on a drifting classifier: the schedule is likely wasted compute.