Retargeted-architecture binary code analysis

Binary code analysis is essential for understanding programs when source code is unavail- able, a common scenario with proprietary software. It underpins critical software security tasks, including plagiarism detection, malware classification, and vulnerability discovery. Security vulnerabilities often transcend architectural boundaries, making cross-architecture analysis crucial for detecting and addressing the same vulnerability across binaries compiled for different Instruction Set Architectures (ISAs).Recent advances in binary code analysis have increasingly leveraged Natural Language Processing models to develop embeddings for assembly languages. However, the diversity of ISAs, compiler optimization levels, and the uneven availability of data across architectures present significant challenges. Training deep learning models for cross-architecture tasks requires substantial data, which is often limited or inconsistent across different architectures. This dissertation introduces an innovative approach termed retargeted-architecture bi- nary code analysis to address data scarcity and reduce per-ISA development efforts. It demonstrates that cross-architecture instruction embeddings can effectively transfer knowledge from one ISA to another, enabling a deep learning model trained on one architecture to generalize across others without retraining. The proposed approach tackles the challenges of cross-architecture instruction embeddings, focusing on both low-level assembly languages and intermediate representations. The results show strong performance in multi-ISA bi- nary analysis tasks, highlighting effective cross-architecture knowledge transfer and robust generalization across diverse ISAs. This work advances the field of binary code analysis by addressing key challenges in cross- architecture contexts, laying the groundwork for future research and practical applications in secure and efficient software analysis.

Author

• Wang, Junzhe

Identifier

hdl:1920/14818

OAI identifier oai:identifier

oai:MARS:1920/14818