Has My Code Been Stolen for Model Training? A Naturalness Based Approach to Code Contamination Detection

It is often valuable to know whether a given piece of source code has or hasn’t been used to train a given deep learning model. On one side, it helps avoid data contamination problems that may exaggerate the performance of evaluated models. Conversely, it facilitates copyright protection by identifying private or protected code leveraged for model training without permission. To this end, automated approaches have been proposed for the detection, known as data contamination detection. Such approaches often heavily rely on the confidence of the involved models, assuming that the models should be more confident in handling contaminated data than cleaned data. However, such approaches do not consider the nature of the given data item, i.e., how difficult it is to predict the given item. Consequently, difficult-to-predict contaminated data and easy-to-predict cleaned data are often misclassified. As an initial attempt to solve this problem, this paper presents a naturalness-based approach, called Natural-DaCoDe , for code-completion models to distinguish contaminated source code from cleaned ones. Natural-DaCoDe leverages code naturalness to quantitatively measure the difficulty of a given source code for code-completion models. It then trains a classifier to distinguish contaminated source code according to both code naturalness and the performance of the code-completion models on the given source code. We evaluate Natural-DaCoDe with two pre-trained large language models (e.g., ChatGPT and Claude ) and two code-completion models that we trained from scratch for detecting contamination data. Our evaluation results suggest that Natural-DaCoDe substantially outperformed the state-of-the-art approaches in detecting contaminated data, improving the average accuracy by 61.78%. We also evaluate Natural-DaCoDe with method name suggestion task, and it remains more accurate than the state-of-the-art approaches, improving the accuracy by 54.39%. Furthermore, Natural-DaCoDe was tested on a natural language text benchmark, significantly outperforming the state-of-the-art approaches by 22% . It may suggest that Natural-DaCoDe could be applied to various source code related tasks besides code complete.