As a representative of diffusion models, denoising diffusion implicit models (DDIM) enable efficient image generation through a deterministic sampling mechanism. However, the invertible nature of their generative trajectory introduces privacy risks, rendering them vulnerable to malicious model extraction attacks. A DDIM model extraction method based on the deterministic and invertible sampling path is proposed in this paper. By exploiting the determinism and reversibility inherent in the DDIM generation process, intermediate latent states and their corresponding predicted noise are reconstructed from a limited number of queries. A surrogate dataset is then constructed using these reconstructed pairs, facilitating high-fidelity functional replication of the target DDIM. Experimental results demonstrate that the extracted surrogate model, trained with only a small number of queries, achieves high generation fidelity while preserving diversity, thereby revealing a critical security vulnerability in the DDIM architecture.
CINÀA E, GROSSEK, DEMONTISA, et al. Wild patterns reloaded: a survey of machine learning security against training data poisoning[J]. ACM Computing Surveys, 2023,55(13s):No.294.
[6]
TRAMÈRF, ZHANGF, JUELSA, et al. Stealing machine learning models via prediction APIs[C]// Proceedings of the 25th USENIX security symposium. Berkeley, USA:USENIX Association, 2016:601-618.
[7]
SZYLLERS, DUDDUV, GRÖNDAHLT, et al. Good artists copy, great artists steal: model extraction attacks against image translation models[DB/OL]. (2021-04-26)[2026-01-26].
[8]
HUH L, PANGJ. Stealing machine learning models: attacks and countermeasures for generative adversarial networks[C]// Proceedings of the 37th Annual Computer Security Applications Conference. New York, USA: ACM, 2021:1-16.
[9]
LIUS Y. Model extraction attack and defense on deep generative models[J]. Journal of Physics: Conference Series, 2022,2189(1):No.012024.
[10]
CHENTY, GUOW H, GEP Y, et al. Model extraction attacks on text-to-image generative adversarial networks[C]// Proceedings of the 2024 IEEE Cyber Science and Technology Congress. Piscataway, USA: IEEE, 2024:272-279.
[11]
SHAZ Y, HEX L, YUN, et al. Can’t steal? cont-steal!contrastive stealing attacks against image encoders[C]// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2023:16373-16383.
SONGY, ERMONS. Generative modeling by estimating gradients of the data distribution[C]// Proceedings of the 33rd International Conference on Neural Information Processing Systems. New York, USA: Curran Associates Inc., 2019:11918-11930.
[14]
VAHDATA, KREISK, KAUTZJ. Score-based generative modeling in latent space[DB/OL]. (2021-06-10)[2026-01-26].