Detecting audio forgery using deep learning techniques with attention mechanisms on ResNet++

Detecting audio forgery using deep learning techniques with attention mechanisms on ResNet++

연구 분야: Artificial Intelligence

학회: Signal, Image and Video Processing

The emergence of deepfake technology, along with other forms of audio forgeries, raises new concerns for digital security and forensic examination. This paper proposes an effective audio forgery detection framework that utilizes a modified ResNet++ deep learning model with attention mechanisms and advanced preprocessing techniques. The algorithm includes a robust detection method that incorporates multi-scale feature fusion and transformer-based attention to a Mel-spectrogram. The model successfully overcomes challenges posed by intricate forgery patterns and overlapping background noise, demonstrating enhanced performance in noisy conditions. In addition, the inclusion of systematic preprocessing steps increases estimation accuracy, which is critical for real-time processing. The proposed algorithm shows good results, boasting 99.92% training accuracy, 95.45% validation accuracy, 99.52% precision, and a false positive rate of only 1.58%.