DEEPFAKE DETECTION VIA FEATURE-LEVEL FUSION OF CONVOLUTIONAL NEURAL NETWORKS

Abstract

The rapid rise of deepfake technology poses significant challenges, including misinformation and identity fraud. This provokes a growing need in robust detection systems. This project explores a feature fusion approach to deepfake detection by integrating the feature vectors of multiple CNN models. Four base architectures—Xception, DenseNet-121, ResNet, and Mesonet—and their paired combinations were trained on the OpenForensics dataset and evaluated for binary classification of real and fake images. Cross-dataset testing was conducted on 16,433 video frames from FaceForensics++ and CelebDF datasets. The Xception model achieved the highest base model accuracy of 88.3%, while the Xception + DenseNet-121 combination outperformed all configurations with an accuracy of 89.6% and a macro average F1-score of 0.89. The results show that feature-level combination of complementary feature spaces improves detection performance, highlighting the promise in this direction. The directions where improvement can happen in the future include increased computation power, larger dataset sizes, and preventing diminutive returns. This article serves as a groundwork for improving deepfake detection techniques using feature-level fusion and collaborative model architectures.