research / Malware
Evasion-aware Android malware detection
MSc thesis at Edge Hill University: an evasion-aware Android malware detector based on an optimised KNN model. Distinction, top 1%.
Inside the work.
The thesis evaluates supervised machine learning algorithms for classifying adversarial evasion attacks. The workflow covers dataset preparation, feature extraction, baseline training, and hyperparameter optimisation with the Firefly algorithm.
KNN was the most resilient to adversarial evasion, including when trained on static features (permissions, intents, hardware components, and API calls) and dynamic features (system calls). The trained model reached 99.2% precision and 99.5% recall.
A related paper, “Evasion Aware Android Malware Detection System Based on an Optimised KNN Model,” with H. Rafiq, M. Usman, and V. T. Ta, is submitted to IEEE Transactions and under review.