Detecting DeepFakes: A Deep Convolutional Neural Network Approach with Depth Wise Separable Convolutions
This paper presents a deep learning based approach for detecting deepfake images and videos that uses a deep convolutional neural network architecture that involves depth-wise separable convolutions to classify whether an image or video is real or fake.
Abstract
This paper presents a deep learning based approach for detecting deepfake images and videos. With the rise of free and easily accessible software tools, such as GANs, creating deep-fakes has become effortless. However, detecting these deepfakes has proven to be a significant challenge. Our method uses a deep convolutional neural network architecture that involves depth-wise separable convolutions to classify whether an image or video is real or fake. We trained and evaluated our model on the CelebDF V2 dataset, achieving high accuracy rates of 98.8% and 97.4% for image and video classification, respectively. Our work is a significant contribution towards mitigating the spread of deepfakes, and we plan to expand our research to detect AI-generated audio in future work. We also propose the development of an online browser extension to make our detection method accessible to the general public and to integrate it into various social media and messaging platforms to prevent the spread of deepfakes.