Deep Fake Face Detection using Efficient Convolutional Neural Networks

TL;DR: This study proposes an efficient CNN model for deep fake face detection, achieving 96% accuracy on a 140k-image dataset, outperforming other CNN models including DenseNet121, EfficientNetB0, and MobileNet, which reached 97-98% accuracy within 15 epochs.

Abstract: Deep fake content, especially in images and videos, is spreading at an unprecedented rate. Fake content are generated with the help of advanced deep learning (DL) algorithms such as GANs, autoencoders, and variation autoencoders. This fake content spreads misinformation, causing a severe impact on society by degrading the trustworthiness of content on social media. Mitigating the risk of these techniques can be done by utilizing the power of one of the DL models, which is CNN. This research study focuses on analysing various Deep-Fake (DF) identification techniques that are trained on various datasets with a small number of samples. The proposed work demonstrates an efficient CNN model and three other CNN pre-trained models through transfer learning on a large dataset available on Kaggle, consisting of 140k images of faces. The proposed CNN model achieved a high accuracy of 96%, while DenseNet121 reached 97%. Both EfficientNetB0 and MobileNet demonstrated even higher performance, each achieving an accuracy of 98% within 15 epochs.