FGSM Adversarial Attack Detection On Deepfake Videos

Abstract

Our goal in this work is to create robust detection models that will counter the danger of adversarial attacks on deepfake videos. We selected a subset of the FaceForensics++ dataset, consisting of 1600 movies evenly divided into actual and fake categories, in order to overcome computational limitations.For FGSM attack detection, we used two sophisticated deep learning models: ResNet50 and Xception. In order to enhance the models' capacity to identify pertinent characteristics, we preprocessed the dataset prior to training by removing frames and concentrating on facial areas. We also added several intensities of FGSM adversarial attacks to improve the dataset's durability and diversification. The outcomes of our experiments were encouraging. With very little loss, the Xception model demonstrated remarkable performance, attaining high accuracies of 98.85% in training, 98.58% in validation, and 93.75% in testing. Never-theless, the ResNet50 model encountered difficulties, exhibiting reduced training, validation, and testing accuracies of 85.26%, 83.18%, and 82.50%, coupled with increased losses. Overall, by offering useful techniques for identifying misleading content, our research strengthens the authenticity and reliability of videos.