DeepFake Image Detection Using Deep Learning Methods

Abstract

Recent advancements in Generative Artificial Intelligence (AI) have ushered in an era of hyper-realistic deepfake media, posing substantial risks to individuals and society through misinformation and exploitation, particularly targeting vulnerable populations. Identifying these advanced manipulations is vital for digital forensics and protecting against potential harm. This project proposes a method that combines the advanced face-detection classifier VGG19 with convolutional neural networks (CNNs) to detect deepfake images. Utilizing the second-generation deepfake dataset DeeperForensics-1.0 for training, we first aim to evaluate our model on a testing set from the same dataset, then test it on a newer dataset to study its robust detection and generalization capabilities. Furthermore, we will employ hyperparameter optimization techniques to fine-tune the neural network architectures, enhancing detection accuracy. The results outperformed state-of-the-art studies in detection on the same dataset, achieving 100.0% accuracy. Finally, experiments were conducted with multiple models trained on a wide range of datasets. The experiments testing the models on unseen and newer datasets showed that models trained with a broader range of manipulation techniques have better generalization, with a 36% increase in recall. By precisely identifying deepfake content, this research aims to equip cybersecurity professionals with the tools needed to combat deepfake-related cybercrimes, thus reducing the risk of blackmail and protecting vulnerable individuals.