Binary type inference is a critical reverse engineering task supporting many security applications, including vulnerability analysis, binary hardening, forensics, and decompilation. It is a difficult task because source-level type information is often stripped during compilation, leaving only binaries with untyped memory and register accesses. Existing approaches rely on hand-coded type inference rules defined by domain experts, which are brittle and require nontrivial effort to maintain and update. Even though machine learning approaches have shown promise at automatically learning the inference rules, their accuracy is still low, especially for optimized binaries.

We present StateFormer, a new neural architecture that is adept at accurate and robust type inference. StateFormer follows a two-step transfer learning paradigm. In the pretraining step, the model is trained with Generative State Modeling (GSM), a novel task that we design to teach the model to statically approximate execution effects of assembly instructions in both forward and backward directions. In the finetuning step, the pretrained model learns to use its knowledge of operational semantics to infer types.

We evaluate StateFormer's performance on a corpus of 33 popular open-source software projects containing over 1.67 billion variables of different types. The programs are compiled with GCC and LLVM over 4 optimization levels O0-O3, and 3 obfuscation passes based on LLVM. Our model significantly outperforms state-of-the-art ML-based tools by 14.6% in recovering types for both function arguments and variables. Our ablation studies show that GSM improves type inference accuracy by 33%.