Abstract

Deepfakes, the use of artificial intelligence to create fake media content, have become a big concern in recent days due to their potential misuse in various fields, like politics, entertainment, and finance. While deepfakes have primarily been associated with video manipulation, recent advances in audio manipulation techniques have made it possible to create highly realistic audio deepfake. This paper proposes a deep learning based approach for detecting audio deepfakes. Specifically, the paper introduce a novel architecture, called Audio Deepfake Detection Network (ADDN), which combines convolutional neural networks (CNNs) and attention mechanisms to detect deepfakes in audio recordings. Our approach is inspired by recent advances in image deepfake detection, which have shown that attention mechanisms can be effective in identifying subtle differences between real and fake images. This proposed approach has several implications for the field in deepfake detection. First, by combining CNNs and attention mechanisms, ADDN is able to learn both local and global features in audio recordings, making it more robust to various types of audio deepfakes. Second, ADDN is computationally efficient, making it suitable for real-time deepfake detection. The findings shows the promise of leveraging deep learning methodologies for detecting audio deepfakes, indicating significant areas for further exploration and advancement in this domain.