AI-Based Deepfake Audio Detection Technique from Real and Fake Audio Dataset

Abstract

The fast development of Deepfake technologies has created major challenges regarding audio authenticity throughout cybersecurity, along with journalism and individual privacy domains. Numerous studies investigate deepfake images and videos and the field of deepfake audio remains under investigation. Researchers need to study deepfake audio methods further because this field is not fully developed. The researchers attempted to develop an audio detection model for fakes despite encountering difficulties in this particular area. A deep learning-based framework was designed to conduct deepfake audio detection through the use of a Convolutional Neural Network and Long Short-Term Memory (CNN_LSTM) model which boosted detection efficiency. Our research included embedding and performing analysis through data preprocessing followed by classification of the Real and Fake datasets. This dataset is a combination of ASVspoof2021_LA_eval (Logical Access) and Deep-Voice datasets. The proposed model's performance evaluation included the use of confusion matrices and an accuracy graph. The model demonstrates efficient audio originality discrimination capabilities and reveals important audio characteristics suitable for effective classification systems. The detection rates of developed algorithms are analyzed through the evaluation of false positive and negative probabilities. The results provide an effective base for deepfake audio detection systems to achieve 97.0% accuracy in detecting false audio content. This development improves the authenticity assessment of audio materials when considering technological manipulation.