Forge Audio Detection Using Keypoint Features on Mel Spectrograms

Abstract

Audio copy-move-forgery audio is one of the most popular methods in the field of audio forensic. This type of forgery is created by copying one or more audio segments and pasting it in another position within the same audio. In this study, for detection of the audio copy-move forgery, a new method using a keypoint-based scheme on the Mel spectrogram model of audio is presented. Firstly, Mel spectrogram image is generated from the suspicious audio. Then, SURF keypoints are obtained from each RBG color channel of Mel spectrogram image. The obtained keypoints from each channel are matched via feature vectors to reveal whether the audio file is forged or original. Finally, the proposed post-processing step is applied to eliminate possible false matches. In the method, providing sufficient final matched keypoints according to the threshold value of the number of matches which is determined by experimental studies reveals that the audio file is forged. Experimental studies are carried out on publicly available the pitch-based dataset. The performance results prove that the proposed method is more robust against even under post-processing operations like noise addition, filtering operation, and compression operation.