Efficient Deepfake Detection Using Convolutional Neural Network

In recent times, generative AI has advanced quickly, and Generative Adversarial Network (GAN) has been used to create realistic-looking fake multimedia. It is still quite difficult to identify these Artificial Intelligence (AI) generated fake photos. Therefore, the proliferation of fake media incites fear in social groups and can harm a person's or community's reputation by influencing public attitudes and perceptions about them. The Convolution Neural Network (CNN) is a useful weapon for combating deepfakes. While the majority of approaches use image datasets with a few images and pre-trained models to demonstrate the fake detection accuracy, this paper presents an efficient CNN architecture, which is trained and tested on a balanced dataset of over 1,40,000 images namely 140k Real and Fake Faces dataset available on Kaggle, comprising 1,00,000 training images and 40,000 test and validation images combined. The proposed CNN model's learning rate is increased by applying an adam optimizer and sparse-categorical cross entropy. With a validation accuracy of 95.40%, a testing accuracy of 95.18% and an Area Under Curve (AUC) score of 0.98, the specially trained CNN model establishes a new standard for the identification and categorization of deepfake images.