An open set generation image provenance method based on dynamic hypergraphs
By using a dynamic hypergraph-based approach, higher-order relationships of generated images are explicitly characterized, and stable and discriminative source fingerprints of generated models are learned. This solves the robustness and accuracy problems of source tracing of generated images in open set scenarios, and achieves effective identification of unknown models and accurate attribution of known models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
AI Technical Summary
Existing methods for tracing the origin of generated images lack robustness in open-set scenarios, struggle to effectively identify unknown generative models, are susceptible to perturbations, and fail to balance accurate attribution of known models with rejection of unknown models.
A dynamic hypergraph-based approach is adopted to learn stable and discriminative source fingerprints of generative models by explicitly characterizing the high-order relationships between multiple samples and multidimensional features. A dynamic hypergraph structure is constructed using a bidirectional attention mechanism, and combined with hypergraph convolution operations and composite loss function optimization, to achieve high-precision source tracing and open set recognition of generated images.
It achieves high-precision source tracing and effective rejection in open set scenarios, improves robustness to image perturbations, can adapt to different generation models, and has good generalization and recognition capabilities.
Smart Images
Figure CN122135085A_ABST