Deepfake Network Architecture Attribution

With the rapid progress of generation technology, it has become necessary to attribute origin fake images. Existing works on image attribution perform multi-class classification several Generative Adversarial Network (GAN) models and obtain high accuracies. While encouraging, these are restricted model-level attribution, only capable handling images generated by seen with a specific seed, loss dataset, which is limited in real-world scenarios when may be privately trained models. This motivates us ask whether possible source models' architectures even if they finetuned or retrained under different configurations. In this work, we present first study Deepfake Architecture Attribution architecture-level. Based an observation that GAN architecture likely leave globally consistent fingerprints while traces left model weights vary regions, provide simple yet effective solution named DNA-Det for problem. Extensive experiments multiple cross-test setups large-scale dataset demonstrate effectiveness DNA-Det.