Mi-maml: classifying few-shot advanced malware using multi-improved model-agnostic meta-learning

AbstractMalware classification has been successful in utilizing machine learning methods. However, it is limited by the reliance on a large number of high-quality labeled datasets and the issue of overfitting. These limitations hinder the accurate classification of advanced malware with only a few samples available. Meta-learning methods offer a solution by allowing models to quickly adapt to new tasks, even with a small number of samples. However, the effectiveness of meta-learning approaches in malware classification varies due to the diverse nature of malware types. Most meta-learning-based methodologies for malware classification either focus solely on data augmentation or utilize existing neural networks and learning rate schedules to adapt to the meta-learning model. These approaches do not consider the integration of both processes or tailor the neural network and learning rate schedules to the specific task. As a result, the classification performance and generalization capabilities are suboptimal. In this paper, we propose a multi-improved model-agnostic meta-learning (MI-MAML) model that aims to address the challenges encountered in few-shot malware classification. Specif