TY - GEN
T1 - Query-Based Fingerprinting of Text-to-Image Large Language Models
AU - Josyabhatla, Saketh Ram
AU - Son, Junggab
AU - Kim, Daeyoung
N1 - Publisher Copyright: © 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - The rapid adoption of text-to-image large language models (LLMs) has raised critical concerns regarding model attribution, accountability, and misuse, particularly as these models are frequently fine-tuned, modified, and redistributed. Existing attribution approaches, such as watermarking and semantic query-based methods, often fail under black-box settings or degrade when models undergo prompt variation and fine-tuning. In this paper, we propose a robust query-based fingerprinting framework for attributing text-to-image generating LLMs under realistic deployment conditions. The proposed method constructs a diverse set of discriminative prompts and compares images generated before and after model modification using ORB (Oriented FAST and Rotated BRIEF) feature matching. By focusing on low-level visual artifacts and training-induced patterns, the approach captures intrinsic model characteristics that persist across prompt transformations, fine-tuning, and contextual modifications. We evaluate the proposed method on multiple state-of-the-art text-to-image models under three challenging scenarios: input transformation, model fine-tuning, and prompt modification. Experimental results demonstrate that the method achieves approximately 97% attribution accuracy while operating entirely in a black-box setting and without embedding watermarks or accessing model internals. These findings highlight the effectiveness of the proposed approach for scalable and non-invasive model attribution in dynamic generative AI ecosystems.
AB - The rapid adoption of text-to-image large language models (LLMs) has raised critical concerns regarding model attribution, accountability, and misuse, particularly as these models are frequently fine-tuned, modified, and redistributed. Existing attribution approaches, such as watermarking and semantic query-based methods, often fail under black-box settings or degrade when models undergo prompt variation and fine-tuning. In this paper, we propose a robust query-based fingerprinting framework for attributing text-to-image generating LLMs under realistic deployment conditions. The proposed method constructs a diverse set of discriminative prompts and compares images generated before and after model modification using ORB (Oriented FAST and Rotated BRIEF) feature matching. By focusing on low-level visual artifacts and training-induced patterns, the approach captures intrinsic model characteristics that persist across prompt transformations, fine-tuning, and contextual modifications. We evaluate the proposed method on multiple state-of-the-art text-to-image models under three challenging scenarios: input transformation, model fine-tuning, and prompt modification. Experimental results demonstrate that the method achieves approximately 97% attribution accuracy while operating entirely in a black-box setting and without embedding watermarks or accessing model internals. These findings highlight the effectiveness of the proposed approach for scalable and non-invasive model attribution in dynamic generative AI ecosystems.
KW - Model Attribution
KW - Model Security
KW - Query-based Fingerprinting
KW - Text-to-Image Generation
UR - https://www.scopus.com/pages/publications/105047338856
U2 - 10.1109/BigDataSecurity69476.2026.00015
DO - 10.1109/BigDataSecurity69476.2026.00015
M3 - Conference contribution
T3 - Proceedings - 2026 IEEE 12th Conference on Big Data Security on Cloud, BigDataSecurity 2026
SP - 54
EP - 59
BT - Proceedings - 2026 IEEE 12th Conference on Big Data Security on Cloud, BigDataSecurity 2026
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 12th IEEE International Conference on Big Data Security on Cloud, BigDataSecurity 2026
Y2 - 8 May 2026 through 10 May 2026
ER -