Robust Audio Copy-Move Forgery Detection Using Deep Keypoint Features on High-Resolution Texture Images

Abstract

Audio Copy-Move Forgery (ACMF) is a critical security problem that compromises the integrity of digital audio recordings and complicates forensic investigations. Traditional methods based on time-frequency analyses cannot reliably detect the forged parts of the audio, especially in the face of forensic attacks such as noise addition, compression and filtering, as they cannot preserve high-resolution details. In this study, a new attack-resistant ACMF detection method is presented by extracting and matching keypoints from super-resolution texture images using the deep learning based ALIKE method. For this purpose, firstly, the audio signal is converted to a high-resolution spectrogram with a Coupled-PHCA based technique; After denoising and L channel histogram equalization steps, LBP based texture maps are obtained from the spectrogram image. Then, the keypoints and descriptors are extracted by the ALIKE network. The obtained keypoints are matched by the g2NN method and the false matches are eliminated by the proposed filtering algorithm. The matches obtained as a result of the filtering algorithm are projected onto the time axis and the localization of repeated segments in the audio file is precisely determined. The method has been tested on TIMIT-ACMF, Arabic-ACMF and Turkish ACMF datasets; in the non-attack case, it achieved 92.1%, 93.0% and 94.8% accuracy, respectively; the accuracy was maintained above 90% on all datasets under 30 dB Gaussian noise or 64 kbps MP3 compression. The proposed method showed a clear superiority with higher detection rate compared to window-based, VAD-based and spectrogram-based approaches.