Audio Deep Fake Detection Using an Ensemble Learning Approach

الملخص

The rapidly and continuously evolving tools and techniques based on artificial intelligence pose a serious and growing concern, especially regarding the spread of deepfake technologies and their impact on the credibility and authenticity of various audio files. To address this growing challenge and effectively counter these concerns, A hybrid model was specifically designed and developed to detect deepfake voices. This proposed model relies on the latest deep learning techniques, particularly leveraging the power of one-dimensional convolutional neural networks (CNN), in addition to integrating them with a multilayer perceptron (MLP) network and an optimized XGBoost classifier operating within an integrated stacking ensemble. The approach proposed in this model is based on extracting Mel-frequency cepstral coefficients (MFCC) from audio files in the dataset used for training and testing. Due to the lack of a suitable and specialized dataset of Arabic audio from the Libyan dialect publicly available for this specific purpose, an integrated dataset was created and compiled, containing a variety of audio clips, including genuine audio clips, as well as fake audio clips generated using various artificial intelligence techniques. When testing the model and evaluating its performance, the model demonstrated high accuracy across datasets, achieving 96.67% (three seconds), 97.50% (five seconds), and 97.00% (seven seconds), confirming its exceptional ability to detect subtle anomalies and hidden.