Rapid developments in generative adversarial networks and transformer-based generation models have led to deepfake image synthesis becoming a growing threat to the authenticity of digital media across numerous domains, including social media, healthcare imaging and surveillance systems. Numerous deepfake detection solutions have been created but are generally limited to face centric datasets and based on convolutional neural networks (CNNs) which have difficulty with cross-domain distribution shifts. Therefore, to further advance current deepfake image detection capabilities, we present Cross Domain-Detect: A robust deepfake image detection framework based on transformer-driven technologies that allow for effective cross-domain generalization. This solution is designed to use the Vision Transformer, Swin-T, and DeiT architectures as backbones for both spatial feature extraction and image classification. Also, unlike many previous deepfake detection research studies that only support offline model inference, the current research has integrated online model inference into a Flask-based production environment that supports secure upload verification, explainable attention heatmaps, persistent document storage, and automated PDF report generation. The overall framework will also be evaluated through the application of accuracy, precision, recall, F1 score and AUC metrics. The cross domain-detect framework will achieve 98.5% accuracy with balanced class performance and provide technically sound as well as practically functional solutions for the transparent and reliable detection of deepfake images.
Prasad et al. (2026) studied this question.