Design and Implementation of Binary Function Similarity Analysis System Based on Deep Learning

Binary function similarity analysis is vital for tasks like malware detection, vulnerability identification, and software maintenance. Traditional methods - relying on handcrafted features or rigid structures - often fail to handle diverse architectures, compiler optimizations, and obfuscation. This paper presents a deep learning-based approach that learns semantic vector representations of binary functions directly from assembly instructions. A two-phase embedding model transforms instructions into context-aware vectors, and a self-attention network highlights key instructions and structural patterns. Using a Siamese architecture and contrastive learning, the system maps similar functions closer together, improving accuracy and scalability. Experimental results on a large, varied dataset show consistently high performance under diverse conditions, demonstrating the method's robustness and potential for practical applications. The method's robustness against code variations makes it applicable to aerospace systems, such as verifying avionics firmware integrity or detecting tampered flight control modules.