QORA: A Sustainable Framework for Open-World Generative Model Attribution with Quasi-Orthogonal Representation Disentanglement

ICLR 2026 Conference Submission16339 Authors

Keywords: model attribution, generative models, open-set recognition, incremental learning

Abstract: The rapid emergence of new generative models poses significant challenges to static attribution frameworks, which often confidently misattribute images from unknown sources to known ones and struggle to adapt stably to new models. To address these limitations, we propose Quasi-Orthogonal Representation Attribution (QORA), a unified framework for sustainable open-world generative model attribution. QORA consists of two core modules. The Progressive Orthogonal Learning Module (POLM) employs Stiefel manifold optimization to construct a quasi-orthogonal feature space that reduces redundancy while maintaining a stable attribution subspace for open-world settings. The Fingerprint Disentanglement and Enhancement Module (FDEM) leverages classifier-guided attention and multi-auxiliary contrastive learning to disentangle and amplify model-specific fingerprints. To enable continual learning, QORA integrates exemplar replay with feature-similarity-based classifier initialization, achieving lightweight incremental updates for new models while avoiding catastrophic forgetting. Extensive experiments demonstrate that QORA achieves state-of-the-art closed-set accuracy and strong open-set robustness across GAN and diffusion benchmarks, while maintaining stable performance during incremental learning, highlighting its superior scalability and applicability in evolving environments.

Primary Area: generative models

Submission Number: 16339