Neural Nets Can Learn Function Type Signatures From Binaries

Function type signatures are important for binary analysis, but they are not available in COTS binaries. In this paper, we present a new system called EKLAVYA which trains a recurrent neural network to recover function type signatures from disassembled binary code. EKLAVYA assumes no knowledge of the target instruction set semantics to make such inference. More importantly, EKLAVYA results are "explicable": we find by analyzing its model that it auto-learns relationships between instructions, compiler conventions, stack frame setup instructions, use-before-write patterns, and operations relevant to identifying types directly from binaries. In our evaluation on Linux binaries compiled with clang and gcc, for two different architectures (x86 and x64), EKLAVYA exhibits accuracy of around 84% and 81% for function argument count and type recovery tasks respectively. EKLAVYA generalizes well across the compilers tested on two different instruction sets with various optimization levels, without any specialized prior knowledge of the instruction set, compiler or optimization level.