LEARNT: A Neural Machine Translation Framework for Accurate Binary Lifting to High-Level Representation

In binary analysis, performing static analyses on architecture-agnostic intermediate representation is efficient and strongly demanded. Sound and accurate Low-Level Virtual Machine Intermediate Representation (LLVM IR) lifted from binary could make the reuse of dozens of existing analysis programs of the LLVM ecosystem possible. However, current binary lifters lack the resources to improve manually developed lifting rules and develop more of them. This work aims to solve the problem of lifting low-level language to sound high-level Intermediate Representation (IR) as a formal language translation problem, enabling automatic learning of binary lifting. Therefore, we propose a neural machine translation-based binary lifting framework named LEARNT with a parallel corpus generation method leveraging a compiler. The evaluation results show that LEARNT’s average translation accuracy is 93%, which proves that translation rules automatically learned by LEARNT are sound.