Zero-Shot Source Code Author Identification: A Lexicon and Layout Independent Approach

We tackle the challenge of Zero-Shot identification of authors of source code, which can be used with no prior samples of authors outside of the training data. In our approach, a feedforward neural network is first trained on a multi-class classification task. Then, a substantial part of this network is duplicated and reused to compare code samples. We refer to this design as Feedforward Duplicated Resolver (FDR) model. We propose new input features to train this model, called Variable-Independent Nested Bigrams, extracted from the Abstract Syntax Trees of code samples. These features provide robustness against lexical and layout obfuscation attacks frequently used in plagiarism attempts. This approach performs accurately even on code samples from unknown authors, on data obtained from Google Code Jam, an international coding competition platform. For example, for the task of predicting whether a pair of samples from 43 unknown authors have been written by the same person, we obtain an AUC of 0.96 and 0.91 for non-obfuscated and obfuscated code, respectively.