Reproduce “Concept Drift Detection and Adaptive Retraining of Malware Classification Models” on OpenResearch

About this paper

Christofer Washington Berruz Chungata, Martin Jurecek, Katerina Potika, William B. Andreopoulos, Mark Stamp

A study evaluates automated concept drift detection methods, including a novel OCSVM-based approach, for adaptively retraining malware classification models. Findings indicate that OCSVM-driven retraining maintains high accuracy (e.g., MLP at 0.9486) while reducing retraining effort by over 65% compared to periodic retraining.

Key insights

• Adaptive retraining based on concept drift detection can achieve classification accuracies comparable to continuous periodic retraining with substantially reduced computational effort.
• Unsupervised machine learning-based drift detectors like OCSVM can be more efficient in terms of detector training time and equally or more effective than statistical methods such as MMD.
• Pareto Front analysis effectively quantifies the trade-off between classification accuracy and retraining efficiency, aiding in optimal configuration selection.

Summary by alphaXiv.