Android applications can leak user's private information and now a days is a major concern. This paper presents MirrorDroid, a novel dynamic analysis based test framework to detect the leakage of private data by Android applications. Our method modifies the Dalvik Virtual Machine to follow execution of the target application with a parallel (mirror) version, which only differs in the sensitive input data. If all other inputs are kept equal, a variance between program output and its parallel version corresponds to an event of sensitive data disclosure. MirrorDroid was evaluated on an Android malware data set and 50 popular applications from the official Android Application Store. The result shows that the proposed system can detect leakage of sensitive data effectively and efficiently.