Enhanced Fake Image Localization in Social Media using Swin Transformer and EfficientNet Feature Fusion

Abstract: Spurious images on social media result in misinformation, online fraud, and manipulation. Conventional CNN-based techniques fail to precisely localize tampered areas because of poor feature extraction abilities and limited generalization. In this paper, the authors suggest a hybrid deep learning method combining CNN with ResNet-50 for enhanced fake image localization. ResNet-50 improves feature extraction through capturing long-distance dependencies and optimal feature propagation, minimizing loss of information. The innovation of this method is efficient spatial and contextual feature fusion through ResNet50, resulting in better segmentation of fake regions and enhanced adversarial robustness. The model is trained on publicly released datasets, including Kaggle Deepfake Dataset and FaceForensics++, and tested against state-of-the-art localization methods. Experimental results indicate better detection accuracy, better segmentation of tampered regions, and adversarial robustness. This method further improves digital forensics through a robust, effective, and economical means of identifying and locating counterfeit images on social media