DeepFake Image Detection Using ReactNet Based Transfer Learning Classification Model on Celeb DF Dataset

This report studies frame-level deepfake detection on the Celeb-DF (v2) benchmark using transfer learning with ReActNet-MobileNet, a binary neural network (BNN) pre-trained on ImageNet. BNNs binarize weights and activations, which makes them attractive for real-time detection on mobile and edge devices. Videos were split at the video level using the official Celeb-DF test list (518 test videos), so no video appears in both training and testing. Frames were extracted at one frame in every ten with OpenCV, giving about 86,500 training and 20,000 test frames. To fit limited GPU resources, a class-balanced subset of 10,000 training frames (5,000 real, 5,000 fake) and 2,000 test frames (1,000 each) was randomly sampled. The model was trained with Adam (learning rate 0.001, batch size 32) for 5, 15, 30, and 50 epochs, followed by a fine-tuning phase at a learning rate of 0.0001 with early stopping. The 30-epoch model generalized best, reaching 63.7% test accuracy (92.8% training, 79.3% validation). Training for 50 epochs raised validation accuracy to 81.4% but lowered test accuracy to 61.45%, and the fine-tuning phase reduced test accuracy to 54.5%. The results show clear overfitting and a large validation-to-test gap, highlighting the difficulty of frame-level detection on Celeb-DF and motivating temporal modelling, attention mechanisms, and stronger regularization in future work.