9th International Conference on ICT for Sustainable Development

Authors - Devanshi Shah, Rachit Shah, Yash Bhadania, Priteshkumar Prajapati, Parth Shah, Dharmendrasinh Rathod
Abstract - With technology advancing, Deep Fake AI techniques make manipulating faces in images and videos easy, notably through precise swapping to create an authentic appearance. Therefore, this research aims to create fake images using traditional and generative adversarial networks (GAN) methods. Additionally, manually generated datasets are incorporated into existing datasets, and using four deep learning algorithms from Convolution Neural Networks (CNN), they are trained and tested to improve accuracy across different datasets. As per outcomes, the ResNet algorithm results in superior accuracy between 98.93% and 99.38% in all three datasets. It has been integrated into an application developed to verify the authenticity of uploaded images. Furthermore, the research introduces an anti-spoofing technique to verify the authenticity of individuals during video calls. It uses real, publicly available videos and images to create composite videos where the maximum portion of the face remains legitimate and only synchronized lips are AI-generated, making it difficult to differentiate between fake and real video. Hence, the anti-spoofing mechanism analyzes the person’s liveliness when the camera is activated, distinguishing between genuine communication and scripted playback. Nevertheless, this experimental study aims to enhance detection accuracy by training and testing datasets of deepfake images and videos, ultimately improving security performance and robustness. These advancements are integrated into an application, enabling users to verify the genuineness of individuals in both images and video calls.