Binary code similarity analysis (BCSA) is a fundamental building block for various software security, reverse engineering, and re-engineering applications. Existing research has applied deep neural networks (DNNs) to measure the similarity between binary code, following the major breakthrough of DNNs in processing media data like images. Despite the encouraging results of DNN-based BCSA, it is however not widely deployed in the industry due to the instability and the black-box nature of DNNs.

In this work, we first launch an extensive study over the state-of-the-art (SoTA) BCSA tools, and investigate their erroneous predictions from both quantitative and qualitative perspectives. Then, we accordingly design a low-cost and generic framework, namely Binaug, to improve the accuracy of BCSA tools by repairing their input binary codes. Aligned with the typical workflow of DNN-based BCSA, Binaug obtains the sorted top-K results of code similarity, and then re-ranks the results using a set of carefully-designed transformations. Binaug supports both black- and white-box settings, depending on the accessibility of the DNN model internals. Our experimental results show that Binaug can constantly improve performance of the SoTA BCSA tools by an average of 2.38pt and 6.46pt in the black- and the white-box settings. Moreover, with Binaug, we enhance the F1 score of binary software component analysis, an important downstream application of BCSA, by an average of 5.43pt and 7.45pt in the black- and the white-box settings.