A hybrid SuperPoint-based feature extraction and classification approach for audio copy-move forgery detection

Creative Commons License

Altay Ş. Y.

Gumushane Universitesi Fen Bilimleri Dergisi, cilt.16, sa.3, ss.878-898, 2026 (Scopus, TRDizin)

Özet

Copy-move forgery, a widely used and content-based manipulation method in audio forensics, is performed by copying a portion of an audio file and pasting it into a different time period. The development of digital audio editing software has led to an increase in this type of audio forgery, consequently jeopardizing the reliability of digital evidence and posing threats to legal processes, journalistic activities, and other areas. This study presents a hybrid method for detecting audio copy-move forgery. Using the SuperPoint architecture, which extracts both keypoints and descriptor vectors simultaneously from images using a deep convolutional neural network, keypoints and their descriptors were obtained from Mel spectrogram images of audio recordings, and the statistical features obtained from descriptor matching were fed into various machine learning algorithms for classification. To ensure optimal performance, hyperparameters were tuned using Random Search, which revealed Random Forest as the most effective classifier for this task. Experimental results on a Turkish audio dataset show that the proposed method significantly outperforms traditional handcrafted feature-based techniques, achieving superior accuracy and F1-score values.