TruMuzic: A Deep Learning and Data Provenance-Based Approach to Evaluating the Authenticity of Music

Abstract

The digitalization of music has led to increased availability of music globally, and this spread has further raised the possibility of plagiarism. Numerous methods have been proposed to analyze the similarity between two pieces of music. However, these traditional methods are either focused on good processing speed at the expense of accuracy or they are not able to properly identify the correct features and the related feature weights needed for achieving accurate comparison results. Therefore, to overcome these issues, we introduce a novel model for detecting plagiarism between two given pieces of music. The model does this with a focus on the accuracy of the similarity comparison. In this paper, we make the following three contributions. First, we propose the use of provenance data along with musical data to improve the accuracy of the model’s similarity comparison results. Second, we propose a deep learning-based method to classify the similarity level of a given pair of songs. Finally, using linear regression, we find the optimized weights of extracted features following the ground truth data provided by music experts. We used the main dataset, containing 3800 pieces of music, to evaluate the proposed method’s accuracy; we also developed several additional datasets with their own established ground truths. The experimental results show that our method, which we call ‘TruMuzic’, improves the overall accuracy of music similarity comparison by 10% compared to the other state-of-the-art methods from recent literature.