Convolutional Neural Network-Based Detection and Identification of Deepfake Audio

Abstract

This study addresses the pressing concern of deepfake audio detection through a comprehensive approach involving convolutional neural networks (CNNs), audio feature extraction, and model evaluation. The research utilises a dataset from Kaggle with labelled deepfake and real audio files and employs the Librosa library for feature extraction, followed by data organisation and preprocessing. Exploratory Data Analysis (EDA) provides insights into class distribution and statistical summaries. Two models, a neural network and LightGBM, are trained and evaluated, with LightGBM exhibiting superior accuracy, precision, recall, and F1-score. The study's implications emphasise the need for ongoing research on emerging deepfake technologies and recommend deploying the robust LightGBM model in real-world scenarios. The study advances understanding of fake audio detection methods, particularly amid the evolving landscape of deepfake techniques, and underscores the importance of adaptable, accurate models for safeguarding against potential threats posed by manipulated audio content.