Enhancing Malware Detection using Deep Learning Approach
Hayat Hussain Reshi, Karan Singh · 2024
In the face of the rapidly evolving malware landscape, traditional detection methods encounter a formidable challenge due to critical vulnerabilities arising from biased training datasets skewed toward common malware strains. This imbalance compromises adaptability and accuracy, leaving models susceptible to novel variants and mutations and plummeting detection rates. However, a promising solution emerges in the form of Variational Autoencoders (VAEs). These neural networks excel at extracting and compressing low-dimensional representations from complex data. Applied to imbalanced malware datasets, VAEs offer a dual approach by learning robust features through disentangling latent factors within the data, separating noise from genuine malware signatures. Additionally, VAEs facilitate data augmentation by generating synthetic malware samples based on the learned latent space, effectively addressing the imbalance problem. This strategy enriches the training pool, enabling the model to generalize more effectively and enhance detection accuracy for rare or unseen variants. Leveraging VAEs to combat data imbalance holds excellent potential for bolstering the agility and resilience of existing detection methods, providing a promising avenue for enhancing the defensive landscape in the ever-evolving Android malware threat scenario, particularly in mobile security.