Detection of copy move audio forgeries targeting audio files

In today's world, the accuracy and reliability of audio recordings used in communication, judicial processes, and voice-activated systems are of great importance. Audio recordings can be obtained with or without the consent of individuals, and these recordings can be used to create fake audio. The most common and easily performed type of audio forgery is the copy-paste operation. Two methods have been proposed for detecting copy move audio forgery: clustering of key points and classification of graph images derived from audio. In the first method, key points and descriptors were extracted using BRIEF, AKAZE, SIFT, and SURF techniques, and these descriptors were then clustered using the OPTICS, k-Means, and g2NN algorithms. This method achieved an average recall of 0.90 on the TIMIT dataset for no post-processing fake audios and various attack scenarios. In the second method, SIFT key points extracted from a high-resolution spectrogram were clustered using g2NN, and after identifying regions with dense matches, the corresponding frequency range was determined, and the audio was processed with a band-pass filter. By applying the Fast Fourier Transform, the frequency domain was accessed, and frame-by-frame features were extracted using the proposed spiral structure. These features were used to generate graph images, which were then classified to label the audio as either fake or original. This methods show F-scores of 0.95, 0.87, and 0.92 on the Arabic, TIMIT, and Turkish datasets, respectively.