Many AI malware detectors are black boxes: they give a verdict but not a reason, and attackers can trick them. This research builds a detector that explains its decisions and resists adversarial evasion.
The problem
Signature-based detection misses new malware, and black-box machine learning lacks transparency and can be fooled by adversarial changes. Security teams need detectors that are accurate, understandable and robust.
What the research did
We combined a LightGBM ensemble classifier, trained on features of Windows executable files from the EMBER benchmark, with LIME and SHAP explanation tools and adversarial defences such as adversarial training and defensive distillation. We tested robustness against FGSM and PGD attacks.
What we found
The framework achieved 91.16% accuracy and a ROC-AUC of 0.974 with stable explanations. Under adversarial attack, it kept much higher performance than undefended baselines, and supported near real-time detection in a simulated cloud environment.
Who benefits
- Security analysts, who can see why a file was flagged
- Organisations needing auditable AI decisions
- Researchers working on trustworthy AI
Limits and what comes next
The work integrates existing methods rather than inventing a new learning algorithm, and was tested on a benchmark and in simulation.
The team
This research was led by Abdullah Al Siam with Moutaz Alazab, Areej Obeidat, me and Imran Mahmud.
Frequently asked questions
What is explainable AI?
AI that shows the reasons behind its decisions, not just the result.
What are LIME and SHAP?
Tools that explain which features most influenced an AI model’s decision.
Why does robustness matter?
Attackers can tweak malware to fool AI; robust models resist that.