A Novel Knowledge Distillation Framework with Intermediate Loss for Android Malware Detection

With the popularity and self-contained functions of Android operating system, malicious attackers have targeted it primarily. Along with attackers are gradually skilled in avoid deep learning detectors, many in-depth researches on Android malware detection have been done. However, these works have complex models and enormous parameters. To settle this dilemma, we proposed a knowledge distillation architecture with intermediate loss to narrow the capability gap by promoting the student network emulate valuable hint knowledge from the intermediate layers of teacher network, based on Multi-Layer Perceptron (MLP). Besides, our framework utilizes static based features, namely permissions and vulnerabilities to effectively characterize applications and construct dataset. We evaluated our framework from various performance metrics and compared with other state-of-the-art deep neural networks. The experimen-tal results indicates that our framework own better performance and is perspective.