Abstract
As digitization is increasing, threats to our data are also increasing at a faster pace. Generating fake videos does not require any particular type of knowledge, hardware, memory, or any computational device; however, its detection is challenging. Several methods in the past have solved the issue, but computation costs are still high and a highly efficient model has yet to be developed. Therefore, we proposed a new model architecture known as DFN (Deep Fake Network), which has the basic blocks of mobNet, a linear stack of separable convolution, max-pooling layers with Swish as an activation function, and XGBoost as a classifier to detect deepfake videos. The proposed model is more accurate compared to Xception, Efficient Net, and other state-of-the-art models. The DFN performance was tested on a DFDC (Deep Fake Detection Challenge) dataset. The proposed method achieved an accuracy of 93.28% and a precision of 91.03% with this dataset. In addition, training and validation loss was 0.14 and 0.17, respectively. Furthermore, we have taken care of all types of facial manipulations, making the model more robust, generalized, and lightweight, with the ability to detect all types of facial manipulations in videos.