Context-Guided Decompilation: A Step Towards Re-executability

When the original source code for a piece of software is not available, binary decompilation provides an alternative approach for software security analysis, reverse engineering, and understanding malware. The traditional decompilation methods fail to produce source code that can be recompiled and run due to issues related to re-execution and compiler optimizations. The latest developments in large language models (LLMs) have resulted in new neural-based approaches to decompilation. However, the code produced is only semantically plausible and not executable, which means that such approaches will have a less practical use. The primary reason for this limitation is the intricacies of compiler optimizations and the loss of semantics during the compilation

process. Those are the reasons that LLMs cannot reconstruct the code, and they do not have sufficient contextual instructions to do so. To solve these challenges, this paper presents a new hybrid framework for decompilation called ICL4Decomp. This framework implements in-context learning (ICL) to provide specific direction to LLMs so they can produce source code that is re-executable. When ICL4Decomp was examined on various datasets, optimization levels, and compiler combinations, it demonstrated an improvement of 40% in re-executability when compared to the current leading decompilation techniques while sustaining its excellent performance.

Message: Hi FE, I need help with “Context-Guided Decompilation: A Step Towards Re-executability” in “Computer Science & Engineering”