Detection of Image-Based Deepfake using Deep Transfer Learning Algorithms

The proliferation of deepfakes, hyper-realistic synthetic media, poses significant global threats including misinformation, fraud, and identity manipulation. Conventional detection methods struggle to keep pace with their evolving sophistication, leading to generalization issues. This study proposes a robust deepfake detection framework leveraging deep transfer learning. The research utilized a comprehensive deepfake image dataset from Kaggle, comprising 70,000 real faces from Flickr and 70,000 fake faces generated by StyleGAN, each sized 256x256 pixels and meticulously split into training (100,000), validation (20,000), and test (20,000) sets. The data underwent extensive preprocessing, including augmentation (e.g., rotation, flipping) and essential color conversion (RGB to BGR) with ImageNet-centric normalization, crucial for optimizing image quality and compatibility with pre-trained models. Three pre-trained Convolutional Neural Network architectures—Xception, EfficientNet, and ResNet—were comparatively evaluated for their feature extraction and classification capabilities. Performance was rigorously assessed using accuracy, precision, recall, and F1-score. Results indicated Xception demonstrated balanced detection capabilities across real and fake classes (e.g., 88% overall accuracy, 87% F1-score for the fake class). EfficientNet achieved the highest overall accuracy (94.0%), showcasing superior discriminatory power. While ResNet showed comparatively lower accuracy (76.36%), all models benefited from the comprehensive preprocessing. This research highlights the effectiveness of deep transfer learning and meticulous preprocessing in enhancing deepfake detection, contributing significantly to digital media integrity and trust. Future studies should focus on validating these frameworks on larger, more diverse datasets, exploring adversarial attack robustness, and investigating real-time detection capabilities.