Developing Adversary-Aware Classifiers for Android Malware Detection

智能手机在人们的日常生活中变得越来越普遍。然而，据报道，考虑到Android已经占据了移动的手机88%的市场份额，每五个Android应用程序中就有一个实际上是恶意软件。机器学习作为一种有效的技术，已被广泛用于检测Android恶意软件。然而，最近的研究表明，精心制作的恶意软件使机器学习无效。在这个项目中，我们提出开发一系列新技术，如1）Android上下文分析，2）基于包装器的爬山算法，3）集成学习，来解决这个问题。这些成果将帮助澳大利亚获得对抗机器学习和移动的安全方面的尖端技术。

Smartphones have become increasingly ubiquitous in people’s everyday life. However, it was reported that one in every five Android applications were actually malware, considering that Android has taken 88% market share of mobile phones. As an effective technique, machine learning has been widely adopted to detect Android malware. However, recent work suggests that deliberately-crafted malware makes machine learning ineffective. In this project, we propose to develop a series of new techniques, such as 1) Android contextual analysis, 2) wrapper-based hill climbing algorithm, and 3) ensemble learning, to solve this problem. The outcomes will help Australia gain cutting edge technologies in adversarial machine learning and mobile security.