A Hybrid Approach to Machine Learning-Based Deepfake Video Detection

Abstract

In modern times, people regularly share a large number of videos on the internet by posting in different social media or online video sharing platforms. These videos are floating around the internet and are easily accessible to the public. The emergence of generative AI technology has made the process of manipulating these videos very easy. One common method is to replace an individual's face or copy facial features and movements in a video. The entire process is done so smoothly that the final video looks almost real, even though it is fake. This type of video manipulation is known as a deepfake. Furthermore, it creates serious concerns about security, misinformation, and personal privacy, as these videos often portray individuals doing things they never did. However, these circumstances can be tackled by differentiating fake videos from real ones, creating an effective detection system. In this research, a publicly available UADFV deepfake dataset was selected. A hybrid approach incorporating machine learning and its subset technology was proposed to detect fake videos by analyzing both spatial and temporal features present in frame sequences extracted from the videos. The model was created effectively by combining multiple convolutional neural networks for spatial analysis with a bi-directional recurrent neural network understanding the temporal dependencies across video frames. This hybrid structure detects both frame-level visual inconsistencies and unnatural frame transitions that, in most cases, flag the content as deepfake. In addition to the hybrid architecture, different effective preprocessing techniques were applied to clarify the video frames and highlight subtle inconsistencies before model training. It is also worth noting that interpretability in an AI detection system is very important. Thus, the model also incorporated explainable AI methods to illustrate which area in the facial region impacts the model's final binary prediction the most. The interpretability of this model further validates its final decision. Overall, this study aims to identify deepfake videos with high accuracy using an effective hybrid modelling approach, and also involves interpretability in the model's final outcome.