Deepfake Audio Detection Using Machine Learning

Le résumé fourni par la source

The fast growth of generative technologies has made synthetic, or deepfake, audio a significant threat to cybersecurity, digital fraud, and voice authentication. The proposed project is a machine learning method that separates real from artificially generated audio clips. Key audio features are generated through the extraction of acoustic measures using Librosa, as well as well-established measures such as Mel Frequency Cepstral Coefficients (MFCCs), chroma, and spectral contrast, that may allow for subtle changes in the quality of speech to be detected. The proposed system pulls out informative audio features such as Mel-Frequency Cepstral Coefficients (MFCCs), chroma features, and spectral contrast to characterize the unique spectral-temporal patterns of human speech. A Random Forest classifier is then trained with these features to predict whether audio is deepfake or real and to perform binary classification of curated datasets of real and fake audio clips. To increase generalizability and robustness, noise-based augmentation is used during preprocessing. The proposed system achieves an accuracy of 97.8% and an F1-score of 98.7 on the SceneFake dataset. It is deployed using a Streamlit-based interface that allows users to upload their own audio files, view the waveform and spectrogram of the audio, and receive classifications with confidence scores in real-time. This app outputs the results with clear visual indicators. The proposed solution illustrates a lightweight, efficient, and practical overall framework for detecting deepfake audio without the complexity and computational overhead associated with deep learning models and processes, making it a feasible option for both real-time deployment and low-resource environments. The Deepfake Audio Detection System aims to advance digital forensics and AI safety by combating the misuse of synthetic media, while demonstrating the role of machine learning in enhancing security and societal awareness. The key contributions of this work are: (1) a lightweight, feature-based deepfake audio detection pipeline optimized for low resource environments; (2) an empirical evaluation showing competitive performance against more complex deep learning approaches; and (3) a real-time deployable Streamlit-based interface enabling practical forensic usage.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.