Skip to Main Content
Malware has posted a great risk to the privacy of users and data. Unfortunately, current anomaly detection is both coarse-grained and ineffective to detect them, because some behaviors can only be triggered in specific circumstances, moreover, it is difficult to analyze and evaluate compromised data. In this paper, we address the two problems via taint-based analysis. First, we precisely characterize malware behaviors in virtual environments, where behaviors can be completely explored by tainting return values of security-related system calls in order to traverse multiple execution branches. Second, sensitive data are also tainted to track their propagation to determine whether they are transmitted maliciously. A supporting distributed architecture is presented to optimize efficiency. Our approach is an effective complement to current anomaly-based models. The initial experiment demonstrates our system can detect a variety of malware with high accuracy and low overhead.