Feature Augmented Meta-Learning on Domain Generalization for Evolving Malware Classification

Fangwei Wang, Yinhe Chen, Ruixin Song, Qingru Li, Changguang Wang

Abstract: The proliferation of malware has resulted in substantial harm to various sectors and economies. Various deep learning-based malware classification methods have been suggested as a means of mitigating malware threats. These methods typically operate under the assumption of independent and identically distributed training and test data. However, this assumption becomes invalid with the evolving malware family. While domain adaptation models offer a potential solution to this issue, their implementation is hindered by the difficulty of collecting new malware variants. In order to address the previously mentioned problem, we suggest an image-based technique for categorizing malware families utilizing domain generalization. Initially, malware is transformed into gray-scale images that depict byte patterns of the malware. Subsequently, these gray-scale images are fed into a model incorporating convolutional block attention to extract features. Furthermore, data augmentation is implemented at the feature level to broaden the distribution of the source domain and enhance the model’s generalization capabilities. Finally, meta-learning is utilized as a training approach to effectively extract domain-invariant representations. A series of experiments are conducted on the BIG2015 and BenchMFC-G1P1P2. The proposed method demonstrates a higher accuracy rate of 88.66% on the BIG2015 and 80.25% on the BenchMFC-G1P1P2, which is better than the existing methods.

External IDs:doi:10.1007/978-981-96-1528-5_14