Deepfake Detection in Call Recordings: A Deep Learning Solution for Voice Authentication

The emergence of deepfake technology has improved exponentially and this intensified the fears that surround the credibility of audio recordings, in instance telecommunication and security. This project proposes a full deep learning-based approach to deepfake voice recordings detection in call communications as an improvement to the voice authentication processes used. It is with this in mind we developed an adaptive architecture which arrange convolutional neural networks (CNN) and recurrent neural network (RNN) in a manner that helps in discerning between real and fake sounds. The nature of the problem allows the use of a large amount of data collected from a wide variety of real and fake audio samples which serve for proper training and testing of the system. To improve the performance of the model, some strategies have been implemented including audio preprocessing such as spectrogram and features. This research adds to the existing body of literature on voice authentication but also seeks to underscore the need for solutions that secure audio communication in times when deepfakes are on the rise. Subsequent research will be dedicated to perfecting the existing model and assessing the feasibility of its use in practice.