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

Malware 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. Specifically, we propose two data augmentation techniques to improve the classification performance of few-shot malware. These techniques involve utilizing grayscale images and the Lab color space. Additionally, we customize neural network architectures and learning rate schemes based on the representative few-shot classification method, MAML, to further enhance the model's classification performance and generalization ability for the task of few-shot malware classification. The results obtained from multiple few-shot malware datasets demonstrate that MI-MAML outperforms other models in terms of categorical accuracy, precision, and f1-score. Furthermore, we have conducted ablation experiments to validate the effectiveness of each stage of our work.