Codee: A Tensor Embedding Scheme for Binary Code Search

Given a target binary function, the code search retrieves top-K similar functions in repository, and represent that they are compiled from same source codes. Searching is particularly challenging due to large variations of compiler tool-chains options CPU architectures, as well thousands Furthermore, there some pivotal issues current schemes, including inaccurate text-based or token-based analysis, slow graph matching, complex deep learning processes. In this paper, we present an unsupervised tensor embedding scheme, Codee, carry out efficiently accurately at function level. First, use NLP-based neural network generate semantic-aware token embedding. Second, propose efficient basic block generation algorithm based on representation model. We learn both semantic information instructions control flow structural Then all embeddings obtain variable-length feature vector. Third, build singular value decomposition, which compresses vectors into short fixed-length facilitate afterward. further dynamic compression incrementally update database. Finally, local sensitive hash method find top- $K$ matching repository. Compared with state-of-the-art cross-optimization-level such Asm2Vec DeepBinDiff, our scheme achieves higher average accuracy, shorter vectors, faster performance using four datasets, OpenSSL, Coreutils, libgmp libcurl. other cross-platform Gemini, Safe, recall also outperforms others.