Core Concepts
Using Generative Adversarial Networks (GANs) for data augmentation in Android malware detection improves model performance and reduces storage requirements.
Abstract
The study explores using GAN-generated data to train a model for Android malware detection. It proposes a method to synthetically represent data using GANs, enhancing classification models. By comparing real and synthetic images, the research shows improved performance with F1 scores reaching 97.5%. The study highlights the impact of image size, malware obfuscation, and different GAN models on classification effectiveness. Data augmentation through GANs proves beneficial in improving model accuracy for detecting Android malware applications.
Stats
The achieved F1 score reached 97.5%
Over 40,000 apps were acquired for the study spanning three weeks and consuming over 100 gigabytes of storage.
WGAN-generated images showed slightly superior quality compared to DCGAN-generated images.
The FID scores of WGAN-generated datasets reached lower values faster than those of DCGAN-generated datasets.
The classification model trained on WGAN-generated data performed better than that trained on DCGAN-generated data.