Doreen Joseph

The purpose of this work is to advance binary code security testing by developing automated and tailored binary analysis techniques. We leverage Transformer-based machine learning models equipped with both code syntax and structure information to learn semantically-rich representations of functions in binaries. These representations can then be used in further security analysis tasks.

In this work, we explore the application of machine learning and Large Language Models (LLMs) for binary similarity analysis in cybersecurity, with the goal of improving third party library and vulnerability detection. Traditional signature-based detection methods often struggle to identify unknown vulnerabilities, while binary analysis techniques can be limited by their reliance on exact matches. By leveraging the capabilities of a custom Transformer-based LLM, which excel at recognizing complex patterns in textual and structured data, this work proposes a novel approach to analyzing binary files. We fine-tune the model to understand patterns in disassembled code and byte sequences, generating embeddings that capture subtle similarities between functions. These embeddings are then used for binary similarity analysis, allowing for the identification of both known and previously unseen functions based on structural similarities. This ongoing research aims to bridge the gap between machine learning and binary analysis, creating a robust, scalable solution. We conduct experiments on a comprehensive dataset comprising binary code samples from various architectures, bitness, and compiler optimization settings, and compare our work against the state-of-the-art.