A deep learning framework for the defection of source code plagiarism using Siamese network and embedding models

Abstract

Source code plagiarism represents an ongoing problem that threatens academic integrity and intellectual rights. Various research works on detection approaches have been proposed to overcome prolonged manual inspection as it requires laborious efforts and consumes time. These detection approaches can be categorised into four major domains; software engineering, knowledge discovery, shallow parsing and machine learning. Review of the literature revealed that most of the detection approaches had been evaluated based on the commonly referenced and established six-level classification of source code transformations known as the Faidhi and Robinson spectrum, except for the approaches in the machine learning domain. Thus, this research sought to fill the gap in the absence of a machine learning approach that uses embedding models to detect source code plagiarism and evaluated based on the six-level classification. The objectives of this research are threefold; to extract various embedding sequences as similarity features from source codes using embedding models, to train a Siamese network that learns similarity representations from source code embedding sequences, and to develop a deep learning framework that leverages embedding sequences and Siamese network to identify the most accurate detection based on the standard six-level classification of plagiarism activities defined by Faidhi and Robinson. A deep learning framework that utilised a Siamese network and embedding models is proposed to detect deliberate plagiarism in source codes. The proposed framework split source codes into character-based, word-based and token-based sequences to obtain embedding sequences through Word2Vec and fastText models. These embedding sequences were then used as inputs to the Siamese BLSTM network for learning similarity representations. The experimental results showed that the character-based embedding sequences with Word2Vec, Skip Gram and Negative Sampling (W2V-SGNS) approach and the token-based embedding sequences with FastText, Skip Gram and Hierarchical Softmax (FT-SGHS) approach outperformed the other approaches. The detection results were also found to be able to detect up to level five (i.e., semantic equivalents) of the standard classification. However, future experiments will require a larger dataset and fine-tuning of the Siamese network to reduce overfitting and to improve the generalisation of the trained models on plagiarism attacks.