Deep Fake Images and Videos Detection using Deep Learning

Abstract

Deepfake technology has now become an actual menace in the digital media world, as it has the ability to generate highly realistic manipulated media. It poses significant questions regarding misinformation, identity impersonation, and cyber fraud against public personalities like politicians, celebrities, and influencers. Deepfakes are mainly produced by Generative Adversarial Networks (GANs), autoencoders, and Convolutional Neural Networks (CNNs). Even though GANs create synthetic visual data using adversarial training and competition between a discriminator and a generator, autoencoders are utilized to carry out face-swapping and feature extraction tasks. To foresee and deter the possible abuse of this technology, this study introduced a system for detecting deepfakes using a hybrid deep learning method. The system employs the Xception and EfficientNet models for image-based detection and LSTM networks for temporal inconsistency analysis. The FaceForensics++ database, which contains real and manipulated video samples, provides the training and testing base. The image-based detection module has been proven to be 95% accurate, and the video-based module achieved 87%, showcasing robust performance in differentiating real content from spurious manipulations. The model is also deployed on Streamlit to allow for real-time user interaction, thus making it suitable for use in real-world applications in digital forensics and media authentication. This work enhances the credibility of internet information and neutralizes the increasing menace to society posed by AI-generated fakes.