Attribution as Retrieval: Model-Agnostic AI-Generated Image Attribution
Authors: Hongsong Wang, Renxi Cheng, Chaolei Han, Jie Gui
Organizations: School of Computer Science and Engineering, Southeast University, Nanjing 210096, China · Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China · School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China
With the rapid advancement of AIGC technologies, image forensics will encounter unprecedented challenges. Traditional methods are incapable of dealing with increasingly realistic images generated by rapidly evolving image generation techniques. To facilitate the identification of AI-generated images and the attribution of their source models, generative image watermarking and AI-generated image attribution have emerged as key research focuses in recent years. However, existing methods are model-dependent, requiring access to the generative models and lacking generality and scalability to new and unseen generators. To address these limitations, this work presents a new paradigm for AI-generated image attribution by formulating it as an instance retrieval problem instead of a conventional image classification problem. We propose an efficient model-agnostic framework, called Low-bIt-plane-based Deepfake Attribution (LIDA). The input to LIDA is produced by Low-Bit Fingerprint Generation module, while the training involves Unsupervised Pre-Training followed by subsequent Few-Shot Attribution Adaptation. Comprehensive experiments demonstrate that LIDA achieves state-of-the-art performance for both Deepfake detection and image attribution under zero- and few-shot settings. The code is at https://github.com/hongsong-wang/LIDA
Figures & tables
Figure 1: Comparison between generative image watermarking and our retrieval-based AI-generated image attribution. Our framework is concise, versatile, and easily adapted to new image generators.
Figure 2: Pipeline of the proposed model-agnostic framework LIDA for AI-generated image attribution. LIDA treats image attribution as an instance retrieval problem, and uses low-bit-plane-based generative fingerprint as the input. The training stage consists of two consecutive steps: (a) unsupervised pre-training and (b) few-shot attribution adaptation.
Figure 3: Comparison of AI-generated images and low-bit generative fingerprints from different image generators. Generators include Stable Diffusion [ 39 ] , ADM [ 13 ] , and Wukong [ 59 ] .
Figure 4: Feature distribution of images from different sources for (a) RGB images and (b) low-bit generative fingerprints.
Figure 5: Impact of loss weight for (a) AI-generated image attribution and (b) AI-generated image detection.
Figure 6: Robustness analysis and comparison of adaptation time: (a) Robustness to image degradation under Gaussian blur and JPEG compression, where β represents the degree of blurring and image quality. (b) Comparison of the running time for unsupervised pre-training and few-shot attribution adaptation.
Figure 7: Illustration of retrieval-based image attribution. The top-ranked retrieval results from the database are listed, with the correct result highlighted by a red bounding box. The queries are a real image, a BigGAN-generated [ 2 ] image, and a Midjourney-generated [ 33 ] image, respectively.
Figure 8: Illustration of retrieval-based image attribution. The top-ranked retrieval results from the database are listed, with the correct result highlighted by a red bounding box. The queries are a Wukong-generated [ 59 ] image, a SDV4-generated image, and a SDV5-generated [ 40 ] image, respectively.
Figure 9: Illustration of retrieval-based image attribution. The top-ranked retrieval results from the database are listed, with the correct result highlighted by a red bounding box. The queries are a ADM-generated [ 12 ] image, a GLIDE-generated image [ 34 ] , and a VQDM-generated [ 17 ] image, respectively.
Figure 10: Illustration of low-bit generative fingerprints. Each image is divided into multiple patches, and the patch with the highest similarity score is selected for subsequent processing.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
Institute of Automation, Chinese Academy of Sciences, Beijing, China · Beijing Academy of Artificial Intelligence, Beijing, China · University of Chinese Academy of Sciences, Beijing, China +3