A Systematic Review on Methods for Detection of Deep Fake Audio, Video and Image Contents Using Machine Learning

Deep fake knowledge defines the new trend of creating and distributing multimedia content, rendering the verification of such multimedia content a difficult task, demanding high levels of detection technology. This systematic study therefore equips the readers with an overall understanding of the advances in machine learning algorithms that are incorporated in the detection of deep fake audio, video, and image content. The review does an excellent job of breaking down the detection approaches by the media type and the employed machine learning models. The techniques that are being discussed include the combination of different models for improving the reliability of the model- based approaches. The investigation begins with understanding the concept of deep fakes and the difficulties that arise due to limitations in its detection because of the constant developments in generative models. This paper provides a critical analysis of the techniques used in the identification of audio deep fakes with focus on the use of speakers' features and spectrograms. The article explores how several available neural network architectures perform in detecting spatial and temporal discrepancies, errors produced during the generation process, and deep fakes in video and image forms. This paper also describes the common performance indices and benchmark databases employed in assessing the detection algorithms, and how it stresses on the agenda of standardization for making the assessment comparisons easier. This paper also provides the drawbacks of present approach such as there exists a cat and mouse relation between the detection algorithms and generating techniques, the level of complexity in computation and a question mark to the generality of the study. They also pointed out that there are future directions for the research with reference to the combination of the multi-modal analysis and the inclusion of interpretability and explainability in the detection models.