Advances in DeepFake Detection: Leveraging InceptionResNetV2 for Reliable Video Authentication

With the widespread availability of free AI-powered mobile applications, the creation of realistic and deceptive videos known as "DeepFakes" has become increasingly effortless. Detecting the authenticity of such videos poses a formidable challenge due to the scarcity of discernible traces. In this study, we delve into the application of an image processing algorithm, specifically InceptionResNetV2, for deepfake detection in videos. The amalgamation of ResNet and Inception models in InceptionResNetV2 yields superior accuracy, making it an ideal choice for our investigation. Our research entails the development of an Android-based application employing the proposed algorithm, with MIT App Inventor for application creation and Google Colab for the prediction model. Furthermore, we conduct a thorough evaluation of the application's performance in detecting deepfake videos, employing three models-InceptionResNetV2, EfficientNetB0, and ResNet50-across two datasets: DFDC and CelebDF, totalling 400 datasets. Notably, InceptionResNetV2 outperformed the other models, achieving an accuracy of 93.20% for the DFDC dataset and an impressive accuracy of 97.72% for the CelebDF dataset. Conversely, ResNet50 exhibited the lowest accuracy for the DFDC dataset, at a mere 52.28%, while EfficientNetB0 displayed the lowest accuracy for the CelebDF dataset, measuring 30.81%. This journal publication sheds light on the advancements in deepfake detection, focusing on the efficacy of InceptionResNetV2 as a powerful tool for reliable video authentication and providing valuable insights into the comparative performance of different models in tackling this complex issue.