Asteria-Pro: Enhancing Deep Learning-based Binary Code Similarity Detection by Incorporating Domain Knowledge

Abstract

Widespread code reuse allows vulnerabilities to proliferate among a vast variety of firmware. There is an urgent need to detect these vulnerable codes effectively and efficiently. By measuring code similarities, AI-based binary code similarity detection is applied to detecting vulnerable code at scale. Existing studies have proposed various function features to capture the commonality for similarity detection. Nevertheless, the significant code syntactic variability induced by the diversity of IoT hardware architectures diminishes the accuracy of binary code similarity detection. In our earlier study and the tool Asteria , we adopted a Tree-LSTM network to summarize function semantics as function commonality, and the evaluation result indicates an advanced performance. However, it still has utility concerns due to excessive time costs and inadequate precision while searching for large-scale firmware bugs. To this end, we propose a novel deep learning-enhancement architecture by incorporating domain knowledge-based pre-filtration and re-ranking modules, and we develop a prototype named Asteria-Pro based on Asteria . The pre-filtration module eliminates dissimilar functions, thus reducing the subsequent deep learning-model calculations. The re-ranking module boosts the rankings of vulnerable functions among candidates generated by the deep learning model. Our evaluation indicates that the pre-filtration module cuts the calculation time by 96.9%, and the re-ranking module improves MRR and Recall by 23.71% and 36.4%, respectively. By incorporating these modules, Asteria-Pro outperforms existing state-of-the-art approaches in the bug search task by a significant margin. Furthermore, our evaluation shows that embedding baseline methods with pre-filtration and re-ranking modules significantly improves their precision. We conduct a large-scale real-world firmware bug search, and Asteria-Pro manages to detect 1,482 vulnerable functions with a high precision 91.65%.