Title: Multi-relational Instruction Association Graph for Cross-architecture Binary Similarity Comparison

Abstract: Cross-architecture binary similarity comparison is essential in many security applications. Recently, researchers have proposed learning-based approaches to improve comparison performance. They adopted a paradigm of instruction pre-training, individual binary encoding, and distance-based similarity comparison. However, instruction embeddings pre-trained on external code corpus are not universal in diverse real-world applications. And separately encoding cross-architecture binaries will accumulate the semantic gap of instruction sets, limiting the comparison accuracy. This paper proposes a novel cross-architecture binary similarity comparison approach with multi-relational instruction association graph. We associate mono-architecture instruction tokens with context relevance and cross-architecture tokens with potential semantic correlations from different perspectives. Then we exploit the relational graph convolutional network (R-GCN) to perform type-specific graph information propagation. Our approach can bridge the gap in the cross-architecture instruction representation spaces while avoiding the external pre-training workload. We conduct extensive experiments on basic block-level and function-level datasets to prove the superiority of our approach. Furthermore, evaluations on a large-scale real-world IoT malware reuse function collection show that our approach is valuable for identifying malware propagated on IoT devices of various architectures.