VisualTruth: Deepfake Video Detection System Using Deep Learning
Samruddhi S. Bagave , Gauri M. Raut , Surbhi D. Raut , Prof. Kumud Wasnik
Computer Science and Technology Usha Mittal Institute Of Technology Mumbai, India
Abstract—In recent years, the proliferation of free deep learning-based software tools has made it remarkably straight- forward to create highly convincing ”DeepFake” (DF) videos, which often exhibit few traces of manipulation. While ma- nipulations in digital videos have been possible for decades through visual effects, recent advances in deep learning have dramatically heightened the realism and accessibility of creating fake content, popularly referred to as AI-synthesized media or DF. The process of generating DF using AI tools has become relatively simple. However, detecting these deepfakes presents a significant challenge, as training algorithms to spot them is a complex endeavor. To address this challenge, we have developed an innovative approach to DF detection by harnessing the power of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). Our system employs a CNN to extract frame- level features from videos, which are subsequently used to train an RNN. This RNN learns to classify whether a video has undergone manipulation and is capable of identifying the temporal inconsistencies introduced by DF creation tools. Our research is underscored by the evaluation of our system’s perfor- mance against a substantial dataset of fake videos collected from standard sources,demonstrating competitive results achieved with a straightforward architecture. In the current era, characterized by the ubiquity of Artificial Intelligence (AI), the ease with which multimedia data can be manipulated and fabricated has raised profound concerns. The emergence of AI-generated deceptive content, commonly known as Deepfakes, has introduced complex challenges for the academic community. Existing methodologies for Deepfake identification have exhibited limitations in terms of generalization. This research is centered on addressing the pressing issue of Deepfake detection, with a focus on developing an effective mechanism that can operate robustly in real-world contexts.
Keywords - Deeplearning, Machine Learning, Deepfake Detection, LSTM, ResNext.
