Malware Detection with Deep Neural Networks

I trained a classification model to run on Windows computers and detect malware running on the computer using only system properties/statistics and achieved an accuracy of 99%.

Malware detection is usually framed as a classification problem, but the constraint that matters in practice is latency: a detector that has to run on every process cannot afford to be slow.

A custom deep neural network built from ReLU activation layers reached 99.93% accuracy on binary malware detection at 0.07ms of latency — roughly 75x faster than the Random Forest baselines it was measured against.

The dynamic analysis pipeline runs over the CICmalmem2022 dataset, 58,000+ records, with decision trees performing feature selection to identify the process indicators that actually carry signal.

Extending the model to multi-class classification identified specific malware families at 97.29% accuracy, which turns a yes-or-no answer into something closer to a diagnosis.

I wrote the approach up in more detail on Medium.