albertan017/LLM4Decompile
LLM4Decompile
LLM4Decompile is a toolset and framework for binary-to-source code translation. It uses large language models to transform machine code into readable source code and recover the original logic of compiled executables.
The project includes a specialized pipeline for generating synthetic training datasets by converting source code into assembly pairs. It provides a fine-tuning framework to optimize deep learning models on these binary-to-source datasets, increasing the accuracy of code recovery.
The system also features capabilities for refining decompiled pseudo-code. This process focuses on restoring the structural skeleton and variable names of a binary to improve the readability of disassembled logic.
Features
- Machine Code Decompilers - Translates native binary machine code back into readable high-level source code using large language models.
- Large Language Model Fine-Tuning - Adapts pre-trained large language models to the specific task of binary-to-source code recovery.
- LLM Fine-Tuning Toolsets - Provides a toolset for supervised fine-tuning of models on binary datasets to improve decompilation.
- Sequence-to-Sequence Mappings - Implements neural machine translation architectures to map binary disassembly sequences to high-level code.
- Supervised Fine-Tuning - Optimizes pre-trained language models using binary-to-source datasets to improve code recovery.
- Synthetic Dataset Generation - Creates synthetic training pairs by compiling source code into binary assembly.
- Training Data Generation - Synthesizes training datasets by converting source code into assembly for binary analysis models.
- Binary-to-Source Porting - Reconstructs high-level source code logic by translating compiled binary instructions.
- Binary Decompilers - Transforms binary machine code into readable source code to analyze executable logic.
- LLM-Powered Decompilers - Transforms machine code into readable source code using large language models to recover original logic.
- Reverse Engineering Tools - Provides a workflow for analyzing compiled binaries to restore original program logic and structure.
- Pseudo-Code Readability Improvement - Employs a language model to improve the readability and accuracy of existing decompiled pseudo-code.
- Pseudo-Code Refinement - Provides iterative refinement of raw decompilation output to correct syntax and improve variable names.
- Pseudo-Code Refiners - Implements a language model process to polish disassembled code and restore variable names.
- LLM-Based Refinement - Uses language models to polish and improve the readability of pseudo code from traditional decompilers.
- Dataset Generators - Ships a pipeline for converting source code into assembly pairs to train reverse engineering models.
- Identifier and Structure Recovery - Restores the structural skeleton and variable names of a binary to improve readability.
- Reverse Engineering - Uses large language models to assist in binary code decompilation.
