Ex2Vec: Enhancing assembly code semantics with end-to-end execution-aware embeddings

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Binary code similarity detection (BSCD) is crucial for computer security tasks.
• Current deep neural network (DNN) methods often use masked language modeling (MLM), limiting execution semantic capture.
• Existing techniques struggle to fully represent the functional meaning of code instructions.

Purpose of the Study:

• To introduce Ex2vec, an end-to-end encoding method for high-quality, execution-semantic-rich embeddings for BSCD.
• To develop a novel pre-training strategy for learning instruction impact on register states.
• To improve the accuracy and effectiveness of binary code analysis.

Main Methods:

• Ex2vec simulates assembly instruction execution to capture semantic features.
• A novel pre-training strategy focuses on the impact of instructions on register states, not just co-occurrence.
• Principal Component Analysis (PCA) is used to visualize and validate the semantic clustering of instructions.

Main Results:

• Ex2vec generates embeddings rich in execution semantics.
• Functionally similar instructions demonstrably cluster together in the embedding space.
• Ex2vec significantly surpasses existing state-of-the-art methods in BSCD performance.

Conclusions:

• Ex2vec offers a superior approach to BSCD by effectively capturing execution semantics.
• The method achieves state-of-the-art results on large-scale datasets.
• Ex2vec demonstrates high accuracy in real-world vulnerability detection scenarios.