Audio Forgery Detection Method with Mel Spectrogram

Abstract

With the development of technology, making forged digital audio gets easier. Copy-move is a popular way to create audio forgeries. Copy move audio forgery is carried out by copying a word and pasting it on another word in the same audio. It gets harder to detect the forged area in the same audio because of the similar external factors. Also post-processing operators such as noise addition, compression etc are applied to audios by attackers after creating forged audio to make it even harder to detect forged area. This is another situation that makes detection difficult in practice. In this work, we used mel spectrogram representation of the audios. The proposed algorithm to detect forgery consists of four stages: keypoint detection, word-level feature description and matching, phrase-level feature description, and post-processing. The experimental results show that the proposed method is highly robust to attacks and gives higher accuracy than other studies in the literature.