XGBoost classified malware with 98.62% accuracy in a new arXiv preprint from researchers at BRAC University in Dhaka, beating neural networks, support vector machines, and Extra Trees on a benchmark dataset S¹P². That number sounds like a solution to a problem that has haunted antivirus vendors for years. It might be. But whether it survives contact with real-world malware is a question the paper does not answer.

My read: This is the kind of result that looks spectacular in a table and fragile in production. A 98.62% accuracy on a benchmark dataset tells you the model learned the patterns in that dataset. It tells you nothing about whether those patterns match what a malware author will ship next Tuesday. Other recent preprints in this space claim similarly high detection rates, but none has been stress-tested against adversarial evolution. The Flask client-server system is a nice engineering touch, but a prototype is not a product.

Why signature-based antivirus keeps losing

The paper opens with a problem every security team knows too well. Traditional antivirus tools rely on signatures, essentially digital fingerprints of known threats S¹. When a new piece of malware appears, there is no signature for it yet. The system is blind until someone updates the database. That gap between discovery and protection is where most damage happens.

Machine learning offers a different bet. Instead of matching fingerprints, you train a model on the features of what malicious and benign code look like, then let it classify new files based on patterns it has learned. The promise is detecting never-before-seen malware. The risk is that the model learns patterns specific to the training data, not to malware in general.

Four models, one winner

The authors, ABM Adnan Azmee, Orko Dutta, and Pranto Protim Choudhury from BRAC University's Computer Science and Engineering department P², ran four algorithms against a benchmark malware dataset S¹:

  • Artificial Neural Networks (ANN): the deep learning default
  • Support Vector Machines (SVM): a classic classifier that draws decision boundaries
  • XGBoost: a gradient-boosted decision tree algorithm popular in data science competitions
  • Extra Trees Classifier: an ensemble method that builds many randomized decision trees

XGBoost came out on top at 98.62% accuracy S¹. The paper does not report the individual accuracy figures for the other three models in its abstract, so the margin of victory is unclear from the available evidence.

XGBoost's strength here fits a broader pattern. When features are structured and the dataset is moderate in size, gradient-boosted trees often beat deep neural networks. They are faster to train, easier to interpret, and less prone to overfitting on smaller datasets. Malware feature data, where each file is represented as a vector of attributes, is exactly the kind of tabular input where XGBoost tends to shine.

The Flask system that classifies files in real time

Beyond the benchmark, the authors built a working client-server detection system using Flask, the Python web framework S¹. The setup takes executable files, extracts their features, and classifies them as malicious or benign through the trained XGBoost model. This is a prototype, not a deployed product. The paper describes it as a proof of concept that the model can work in a real-time pipeline.

What to do about it

For a security operations team at a mid-sized financial firm, this paper is a signal, not a solution. If you are running a Security Operations Centre that processes thousands of executable files daily, the idea of supplementing signature-based scanning with an ML classifier is worth piloting. But you would not deploy a model trained on a public benchmark dataset against live threats without first testing it against your own traffic.

The practical step: take the paper's methodology as a template. Train XGBoost on your own labelled malware and benign file samples, validate it on a holdout set drawn from recent traffic, and measure the false positive rate. A 98.62% accuracy means 1.38% of files are misclassified. In a pipeline processing 100,000 files a day, that is 1,380 wrong answers. Whether those are false alarms or missed threats determines whether the system is useful or dangerous.

What we don't know yet

The 98.62% figure is self-reported and has not been independently verified S¹. The paper is a preprint, not peer-reviewed S¹. We do not know the accuracy scores of the other three models, the composition of the benchmark dataset, or how the training and test splits were constructed. The q-fin.GN cross-listing, which places the paper in a quantitative finance category, is not explained in the abstract, and we should not assume a finance-specific focus without reading the full text.

Most critically, we do not know how the model performs against adversarial malware, code specifically crafted to evade detection. Benchmark datasets are static. Real malware evolves. The gap between benchmark accuracy and real-world detection is the central unsolved problem in ML-based malware detection.

The next signal: whether the authors release their dataset and code for independent replication, or whether a revised version appears on arXiv. We will check the arXiv listing for a v2 update before the end of September 2026.

Subscribe to get the next preprint we break down, before it reaches your security team's inbox.

Sources: S1 — Performance analysis of Machine learning algorithms for predicting mal · P2 — Performance analysis of Machine learning algorithms for predicting mal · P3 — PerforatedAI/PerforatedAI · P4 — umich-sota/TF-as-SVM · P5 — jwmng/extratrees

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.