Few-Shot Learning Based on CCGAN-CNN in Android Malware Classification
Jinpei Yan, Lu Gang Yang, Yanze Yi · 2024
Deep neural network has always been the preferred training model for android malicious feature learning with large amounts of data. The number of samples plays a crucial role in the effect of malware classification by deep learning. Recent researches mainly use large and complex deep neural network models (e.g., LSTM or Transformer) to classify the malware, which heavily relying on the original dataset size. In this paper, we propose a novel android malware classification method that learns feature automatically only by few raw samples. Firstly, we generate the grayscale images from the android .apk files, then the proposed model uses CCGAN network to increase malware samples diversity through data augmentation, and combines the improved CNN network for android malware image classification. We make experiments on 40 raw samples, which includes 20 benign software collected from android apps official markets and 20 malwares downloaded by Grebin. The experiment results show that our method achieves 99.65% accuracy for android malware classification. Moreover, we take the android malware classification evaluation on large number of samples, which our model achieves a considerable effect on classification accuracy comparing with related work on a large number of samples.