Malceiver: Perceiver with Hierarchical and Multi-modal Features for Android Malware Detection

We propose the Malceiver, a hierarchical Perceiver model for Android malware detection that makes use of multi-modal features. The primary inputs are the opcode sequence and the requested permissions of a given Android APK file. To reach a malware classification decision the model combines hierarchical features extracted from the opcode sequence together with the requested permissions. The model's architecture is based on the Perceiver/PerceiverIO which allows for very long opcode sequences to be processed efficiently. Our proposed model can be easily extended to use multi-modal features. We show experimentally that this model outperforms a conventional CNN architecture for opcode sequence based malware detection. We then show that using additional modalities improves performance. Our proposed architecture opens new avenues for the use of Transformer-style networks in malware research.