Generative Trace Attribution Network

Abstract: A deepfake is a digitally created or altered image, video, or audio made using artificial intelligence (AI) that seems real but is intended to deceive or mislead viewers. The rapid rise of generative AI tools like GANs and diffusion models has made it easier to create deepfake images. While traditional methods can detect deepfakes, they often fail to identify which model created them. The process of matching a deepfake image to its generative model is known as deepfake attribution. Deepfake attribution is essential for accountability and preventing misuse of generative models in the creation of deefake. Unfortunately, most existing attribution methods only work well on the specific models they were trained on. They struggle to attribute images generated from unseen generators with different initialization seeds, trained for additional epochs, fine-tuned, retrained, or having slight modifications in loss functions or model architecture. To address these limitations, we propose the Generative Trace Attribution Network (GTA-Net), a generalized attribution network that robustly attributes fake images across diverse generative models with variations, including entirely unseen generative models. GTA-Net works by analyzing hidden patterns in input images using a combination of frequency analysis and latent space analysis to capture training-induced artifacts of target generative models to be attributed. GTA-Net also employs supervised contrastive learning to separate features between different target generative models. Extensive experiments on diverse generative models demonstrate that GTA-Net significantly outperforms existing attribution techniques, offering a more robust and reliable approach for deepfake attribution.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Jaesik_Park3

Submission Number: 7864