A Pattern-based Comparative Analysis and Hybrid Model for Deepfake Image Source Attribution

The explosive growth of deepfake-generation software has made it a pressing necessity to have systems that not only identify deepfakes but also correctly classify their generative source—a field that is still underexplored. This paper tackles deepfake source classification by introducing DeepNoiseNet, a dual-stream hybrid model that combines semantic deep features and forensic noise traces. The model combines EfficientNet and Vision Transformer embeddings with Photo-Response Non-Uniformity (PRNU) and sensor residuals in an attention-based mechanism. The architecture is understandable, resilient against compression artifacts, and generalizes well to twelve generative models. State-of-the-art results in experiments include 93.2% accuracy on classification and 0.91 F1-score with excellent resistance to JPEG compression and Gaussian noise. The suggested system is of great potential for real-world digital forensics, helping law enforcement agencies and content verification websites trace tampered media reliably.