| Popis: |
With the large number of Android applications being released, reliable Android malware classifier is required. In recent years, machine learning as well as deep learning have played important roles in Android malware detection, and various static and dynamic features have been extracted and combined with both machine learning algorithms and deep learning algorithms. Studies have shown that these models can detect malware that is not included in the training set, indicating potential abilities to capture zero-day malware samples. However, the use of obfuscation technology, which is originally aimed at optimizing code and protecting intellectual property, has become a kind of threat in this type of detection method because both static features and dynamic features might be influenced by obfuscation, which has a negative impact on efficacy of models based on these features. As a result, the trained models might be overfit. In this work, we propose a deep learning based Android malware detection method that uses obfuscation labels in training to force the deep learning model to learn features of obfuscation technologies and malware simultaneously from part of the input. The experimental results demonstrate that our method achieved 96.2%-99.6% accuracy on different datasets and suppressed overfitting compared to method that doesn’t use obfuscation labels. |
