Deepfake Image Detection: Comparison of Different Convolutional Neural Networks for Image Detection

Abstract

Deepfake technology relies on advanced theoretical concepts of machine learning and deep learning. Machine-generated fake images can present fictional content as real, often by reenacting or swapping faces using artificial neural networks. The human eye cannot easily distinguish between real and fake faces created by deepfake technology. Therefore, people face numerous difficulties due to the lack of an accurate deepfake detection method. The challenge here is to implement a reliable method for detection of deepfakes due to their rapid spread and the ease of generating deepfakes. In this paper, our objective was to analyze the Convolutional Neural Networks (CNNs) based- existing deepfake detection methods to classify input images as fake (0) or real (1). For our experiment, we employed three different convolutional neural network architectures: 1) 2D CNN, 2) DenseNet-121, and 3) VGGFace-16. We compared their performance and further explored the use of grayscale images and data augmentation techniques with the DenseNet-121 architecture. We used Accuracy, F1-score, Recall, and Average Precision metrics to assess our models and ensure the effectiveness of their performance. We used the two most known publicly available datasets (“140K Real and Fake Faces” dataset and “Real and Fake Face detection” dataset), containing both real and synthetic face images. Our study concluded that the implemented DenseNet-121 architecture showed the best performance with an accuracy of 0.96 and F1-score of 0.96. Although the VGGFace-16 model demonstrated comparable accuracy (0.89), it is computationally very expensive and requires a sophisticated processor in training augmented data. The performance of the custom CNN model (accuracy - 0.90) was comparable to that achieved by the VGGFace-16 model. Our study indicates that existing CNN- based deepfake detection methods may no longer be fully effective due to advancements in current deepfake datasets.
Keywords: Machine learning, Deepfakes, Convolutional neural networks, Deep learning, Deepfake image detection