Deepfake face image source tracing via latent feature analysis

In forensic identification of deepfake face images, source tracing of specific deepfake methods is critical to extending digital evidence's evidentiary value, yet current research focuses on binary authenticity classification that fails to meet judicial tracing demands. To address this gap, we propose a deepfake face image source tracing method via latent feature analysis: we constructed a 31-category dataset (1 real-face, 30 deepfake classes, 620,000 images) covering mainstream forgery mechanisms, designed an EfficientNet-B4-based multi-classification network with hard-class weighting to extract discriminative latent features, performed hierarchical clustering on latent feature-derived 31-dimensional confidence vectors to determine 10 optimal clusters, and used cluster centers as tracing anchors. Experiments show the base+hard model achieves 0.9568 accuracy, nearly all classes achieve Top-3 accuracy above 0.97, overall tracing accuracy reaches 0.9911, and all 30 forgery classes have homologous consistency rates exceeding 0.95;Validation using the Roop face-swap tool demonstrates preliminary feasibility of the proposed framework for one type of unseen forgery. This work upgrades detection from binary authenticity discrimination to fine-grained source-tracing prototype, providing technical foundations for potential applications in digital forensic identification and criminal investigation. Future work will focus on dataset expansion and real-world adaptability optimization prior to practical forensic deployment.