Semantic-aware framework for zero-shot malware classification <i>via</i> attention-based relation network

Deep neural networks have proven effective in identifying known malware; however, they face challenges when it comes to detecting novel malware that they have not encountered before. This issue arises from their dependence on labeled data for training, which is often scarce for new or uncommon malware types. As a result, creating a model that can detect every possible form of malware becomes impractical. Identifying previously unseen malware is essential, which calls for innovative methods such as Zero-Shot Learning (ZSL). ZSL involves classifying categories that were not present during training. To address this, we propose a novel technique called the Semantic-aware Multi-level Attention-based Relation Network (SMART) for zero-shot malware detection. SMART incorporates Relation-wise Attention (RwA) and Pairwise Semantic Attention (PwA) mechanisms to improve detection accuracy. The PwA component is designed to capture relationships between pairs of input elements, while the RwA mechanism operates at a higher level, analyzing interactions among multiple elements. Our approach outperformed previous methods by significantly reducing false positives and achieving a notable accuracy rate of 95%.