Detecting Deepfake Audio Using Spectrogram-Based Machine Learning Approaches

Deepfake audio has emerged as a significant threat in the digital age, with the potential to deceive listeners and cause harm in various sectors, such as media, security, and social networks. This project focuses on detecting deepfake audio using spectrogram-based machine learning approaches. The system leverages four powerful machine learning algorithms: Convolutional Neural Networks (CNN), EfficientNet, CNN-GRU, and a hybrid CNN-GRU model. These models are applied to spectrogram representations of audio signals, which convert time-domain audio data into visual forms, enabling the models to detect subtle differences between real and fake audio. The CNN and EfficientNet models are utilized to extract features from the spectrograms, while the CNN-GRU hybrid model combines the strengths of CNN for feature extraction and GRU for temporal sequence learning. The project aims to classify audio as either Fake or Real, providing a robust framework for detecting manipulated audio content. The models are trained and evaluated using Python and libraries such as TensorFlow and Keras for deep learning, with performance metrics including accuracy, precision, recall, and F1-score. The goal of this project is to offer a scalable, real-time solution for detecting deepfake audio, addressing the increasing concern over misinformation and malicious use of AI-generated content. By combining deep learning techniques and advanced feature extraction methods, this project emphasizes the growing need for robust detection systems in combating digital audio manipulation.

Keywords: Deepfake Audio, Spectrogram, Machine Learning, CNN, EfficientNet, CNN-GRU, Fake, Real, Audio Manipulation, Feature Extraction, Deep Learning, Real-Time Detection, Python, TensorFlow, Keras.