A Deep Learning Method for Obfuscated Android Malware Detection

With smartphones becoming increasingly prevalent and irreplaceable parts of modern society and individual life, it opens new windows for cybercriminals into our everyday life. Mobile devices are now at risk of being attacked by adversaries ranging from small-time crooks hoping to make a quick profit scamming people to authoritarian nation-states trying to attack and target journalists and human rights activists Malware programs are codes that are written by these adversaries to perform malicious tasks on the victim’s devices. These tasks can be spying on victims, encrypting all their files and extorting money, or using the computational power of a victim‘s device for illegal purposes. Many anti-malware solutions that exist for Android Smartphones rely on manual analysis and signatures to detect malware. However, most modern malware relies on polymorphic tendencies to change the signatures. Another major way of detecting malware is by using Machine Learning and Deep Learning models for detection and classification. One of the major issues faced by most detectors is Adversarial evasion, where an adversary modifies the malware using various techniques to intentionally induce a misclassification by the detectors to evade it. Another major issue in this area is the selection of features from a large pool. Selecting a large number of features and not selecting enough can both be detrimental to the model’s performance. In this project, we propose a novel robust malware detection mechanism based on Deep Learning for Android devices that solves both of the aforementioned challenges and thus creating a secure Android ecosystem.