Harel Berger - Robust Android Malware Detection

A growing number of malware detection methods are heavily based on Machine Learning (ML) and Deep Learning techniques. However, these classifiers are often vulnerable to evasion attacks, in which an adversary manipulates a malicious instance from being detected.
This study offers a framework that enhances the effectiveness of ML-based malware detection systems in the field of Android application packages (APK). This work follows previous work on the PDF domain. This framework analyzes different aspects of defenses based on retraining methods of problem-space and feature-space evasion attacks. Also, several key insights were drawn during this research. The first insight is the creation of the malicious predictor system that tries to predict if an evasion attack is successful. The second insight is the effect of merging two types of feature sets to address evasion attacks of multiple types.