Few-Shot Learning to Classify Android Malwares

Mobile phones have become a target for cybercrime where malicious apps are developed to acquire sensitive information or corrupt data. To mitigate this issue and to improve the security in mobile devices, many machine learning methods have been developed to detect and classify Android malware. One problem with the existing methods is that they all assume a large dataset is available to train the learning models. In this paper, a model named few-shot malware classification (FSMC) is proposed to classify Android apps by using only a few training cases for each class. For three small datasets, the experimental results demonstrate that the FSMC model is able to achieve significantly higher accuracy compared to the existing 2-way 1-shot and 2-way 3-shot malware classification methods.