Structural-semantic embedding with position-aware attention and adaptive contrast for binary code similarity detection

Binary code similarity detection (BCSD) is a critical task for identifying vulnerable components in compiled binaries. However, current methods fail to reliably handle the significant compilation variations across platforms, compilers, and optimization levels. To overcome these challenges, we propose structural-semantic embedding with position-aware attention and adaptive contrast (SPAC), a novel BCSD framework that learns robust structural-semantic embeddings. SPAC integrates a learnable residual graph neural network for deep structural and semantic learning, a position-aware attention module for capturing spatial relationships, and an adaptive contrast aggregator for enhanced robustness and discriminability. This combined design enables SPAC to effectively model function semantics invariant to compilation differences. Evaluated on three large-scale datasets, SPAC demonstrates superior capabilities. Numerous experiments have shown that our method significantly outperforms the current state-of-the-art solutions. For example, in the XM retrieval experiment with a sample pool of 8000, compared with Trex, GMN, and GGSNN, our method achieved improvements of 155.1%, 92.3%, and 119.3% in recall score, respectively.