- Conference Article
- 10.1145/3723890.3723908
Binary Similarity Detection Based on Intermediate Representation Graph Fusion
- Jan 17, 2025
- Xiao Li + 4 more +4
Programs have become more complicated and their control flow and data flow difficult to examine as code obfuscation methods have become widely adopted. Reverse engineering and the identification of obscured functions are thus greatly challenged. While concurrently hiding data dependencies and changing data flow paths, obfuscation methods drastically distort a program's control flow by rearranging code and adding false paths. Analyzing Control Flow Graphs (CFGs) and Data Flow Graphs (DFGs) suffers when current methods fail to precisely capture the true paths of data flow and struggle to restore the original execution logic. Conventional single-graph analysis techniques are insufficient to provide a thorough knowledge and analysis of program behavior, therefore restricting the capacity of analysis tools or algorithms to properly comprehend the running process of a program. Researchers have developed graph fusion methods to overcome this restriction and offer a more complete picture of program behavior. Nevertheless, current graph fusion techniques usually produce graphs directly from binary code and perform direct fusion of graph topologies. This approach reduces the accuracy of similarity detection by obfuscation, hence losing original semantic links between nodes and edges. To overcome these challenges, this work offers a novel graph structure fusion method grounded on LLVM IR instructions. The proposed method effectively catches the complicated relationships between the control flow and data flow of a program, creates a semantically richer graph structure, and lessens the effect of semantic obfuscation on analysis results. Ablation experiments confirm how well graph fusion and LLVM IR raise detection accuracy. Furthermore, comparison experiments with contemporary models (like Graphormer, HermesSim, and EvenNet) reveal that, in terms of accuracy, the dataset generated with the proposed approach far surpasses those generated from alternative graph topologies. Particularly, the proposed method manages obfuscated choices very effectively, getting a P@10 approaching 98.4% and maintaining a P@1 above 80% under usual obfuscation situations. This highlights the efficiency and dependability of the proposed strategy. Particularly, the model proves its durability and efficiency by showing notable accuracy gains under structural obfuscation conditions.
Read more