Advancements in Deep Learning Techniques for Robust Detection of DeepFakes

With tremendous growth in artificial intelligence and machine learning, numerous techniques can be used to distort and manipulate the media. Although many of these techniques are used for scientific or academic purposes, their potential harmful or devious activities are very important to address. One such tool of widespread misuse is the DeepFake technology. A user on Reddit coined the term "DeepFake," which originates from Deep Learning (DL) and is used to generate fake images of people. They are typically used in entertainment and education, but it is important to address the above-mentioned misuse cases. Hence, deepfake detection has received much required attention, and many techniques have been proposed to identify its instances, whether in photos or videos. Various techniques have been implemented to detect deepfakes, such as Convolution Neural Network (CNN), Artificial Neural Network (ANN) and other pre- trained models including ResNet50, VGG16, VGG19, InceptionNet, EfficientNet, Transformer, U-Net and Transfer Learning. The results showed that using MobileNet as a pretrained model in transfer learning outperformed other models by achieving 83% accuracy and the least loss (0.19) among all.