From Heuristics to Transformers: A Comprehensive Survey of Type Inference from Stripped Binaries
Abstract
The recovery of high-level type information from stripped binaries—executables devoid of symbol tables and debugging information—is a cornerstone of software reverse engineering, vulnerability analysis, and decompilation. This survey tracks the evolution of binary type inference from early rule-based heuristics and static analysis to modern deep learning architectures. We analyze the shift from "duck typing" and constraint-solving techniques (e.g., BITY, BinSub) to context-aware neural models (e.g., EKLAVYA, CATI) and finally to state-of-the-art Transformer and Graph Neural Network (GNN) architectures (e.g., SeeType, TYGR). We identify core challenges, including optimization-induced semantics loss and structural type recovery, and propose future research directions in neuro-symbolic inference.