Скачать или смотреть ContraBin: Pre-Training Representations of Binary Code Using Contrastive Learning

This is the recorded presentation of *ContraBin: Pre-Training Representations of Binary Code Using Contrastive Learning*, accepted to TMLR 2025.

Authors:
Yifan Zhang†, Chen Huang‡, Yueke Zhang†, Huajie Shao‖, Yu Huang†, Kevin Leach†
† Vanderbilt University
‡ National University of Singapore
‖ College of William & Mary

---

*Abstract:*
ContraBin is a contrastive learning framework for binary code representation that leverages compiler intermediate representations (LLVM IR) and program analysis techniques to improve binary understanding. It learns robust embeddings by interpolating and contrasting multiple views of source code, compiled binaries, and generated comments. ContraBin achieves significant improvements across diverse downstream tasks, including binary summarization, binary translation, function classification, and function name recovery.

*Approach:*
📐 Introduces a Simplex Interpolation Module that interpolates embeddings between similar binary representations, mimicking human-like learning.
🔄 Pre-trains jointly on source code, compiled binaries, and natural language comments to capture multi-modal program semantics.
🧠 Leverages contrastive alignment to ensure binary embeddings are both discriminative and semantically faithful.

*Key Results:*
🚀 Outperforms strong baselines (RoBERTa, CodeBERT, GraphCodeBERT) on 4 downstream tasks, including classification and summarization.
📊 Demonstrates improved generalization and robustness in function name recovery and binary translation.
🔒 Provides a scalable foundation for LLM-based binary code intelligence and analysis.

---

🙏 Thank you for watching! If you have any questions, feel free to reach out:

Contacts:
Yifan Zhang ([email protected])