Generated Distributions Are All You Need for Membership Inference Attacks Against Generative Models

Minxing Zhang, Ning Yu, Rui Wen, Michael Backes, Yang Zhang

Keywords: generative models, diffusion models, membership inference

TL;DR: Our work proposes a generalized membership inference against various generative models.

Abstract: Generative models have shown their promising performance on various real-world tasks, which, at the same time, introduce the threat of leaking private information of their training data. Several membership inference attacks against generative models have been proposed in recent years and exhibit their effectiveness in different settings. However, these attacks all suffer from their own limitations and cannot be generalized to all generative models under all scenarios. In this paper, we propose the first generalized membership inference attack for generative models, which can be utilized to quantitatively evaluate the privacy leakage of various existing generative models. Compared with previous works, our attack has three main advantages, i.e., (i) only requires black-box access to the target model, (ii) is computationally efficient, and (iii) can be generalized to numerous generative models. Extensive experiments show that various existing generative models in a variety of applications are vulnerable to our attack. For example, our attack could achieve the AUC of 0.997 (0.997) and 0.998 (0.999) against the generative model of DDPM (DDIM) on the CelebA and CIFAR-10 datasets. These results demonstrate that private information can be effectively exploited by attackers in an efficient way, which calls on the community to be aware of privacy threats when designing generative models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/generated-distributions-are-all-you-need-for/code)