INTERNATIONAL RESEARCH JOURNAL OF MODERNIZATION IN ENGINEERING TECHNOLOGY AND SCIENCE

Abstract

The rapid advancement of deep learning technologies has facilitated the creation of convincing face swaps in videos, commonly known as "Deep Fakes" (DF). In recent months, the accessibility of free deep learning-based tools has significantly contributed to the proliferation of this technology. Despite visual effects being employed for decades to manipulate digital video content, recent developments in deep learning have notably enhanced the accessibility and realism of deceptive content.The increased computational power of deep learning algorithms, particularly in the realm of deep fakes, has made the creation of AI-generated media more attainable. However, this accessibility has raised concerns due to the potential misuse of realistic face-swapped deep fakes. These malicious activities include inciting political unrest, orchestrating terrorist acts, spreading revenge porn, and extorting money.In response to these challenges, a novel deep learning technique has been introduced to effectively distinguish between authentic and AI-generated videos. This approach utilizes a Rest Next Convolutional Neural Network (CNN) to extract frame-level characteristics. Subsequently, these characteristics are employed to train a Long Short-Term Memory (LSTM)-based Recurrent Neural Network (RNN). This innovative system can automatically detect deep fakes involving replacement and reenactment, ensuring practicality in real-world scenarios and enhancing model performance on real-time data.The evaluation of this approach involved using a comprehensive and meticulously curated dataset. Various well-established datasets, including the Deep Fake Detection Challenges, Face-Forensic++, celeb-DF, and other pre-existing datasets, were combined to create a robust dataset. This thorough compilation ensures a diverse and representative set of examples for training and testing the deep learning model, enhancing its ability to accurately detect deep fakes across different contexts and scenarios.

Paper File to download :