Exploring Deepfake Detection: A Comparative Study of CNN Models
Raveena, Rita Rana Chhikara, Pooja Punyani · 2024
The rapid advancement of deep learning methodologies has given rise to worries regarding the misuse of hyper-realistic multimedia due to the introduction of deepfake content created by generative adversarial network (GAN) models. Deepfakes, which include altered audio and/or video clips that are nearly identical to real ones, can be used maliciously for things like propaganda, cybercrimes, and political campaigns. To address this challenge, a comparison is conducted involving several CNN models, like EfficientNetB0, VGG-16, DenseNet121, VGG-19, MobileNetV2, ResNet50, InceptionV3, and Xception for deepfake detection. The models are trained using transfer learning technique and by fine-tuning them on the dataset using various hyperparameters. The performance analysis was performed on six cases in which the optimizer, learning rate, batch size, and epochs were adjusted. By exploring this comparative study, a contribution is made to the development of more robust solutions for detecting deepfakes. A thorough analysis of different pre-trained models is conducted and verified, based on the reported outcomes, ResNet50 outperforms the other models. The evaluation of the model's performance involves the comparison of various metrics that have been identified, such as Accuracy, Precision, AUC-ROC curve, and F1-score.