扩散模型的主要思想是在前向过程添加高斯噪音,使得图片逐步逼近标准正态分布,然后训练模型学习逆过程来逐步去噪。对于无条件生成的扩散模型,如 DDPM
[9]
,其损失函数通过优化对数似然的证据下界(
E
vidence
L
ower
Bo
und,ELBO)来得到:
而对于条件生成的扩散模型,如 Stable Diffusion
[10]
,其损失函数通过优化相应的条件对数似然的 ELBO 来得到:
除了训练过程,在后文中,本文也通过 ELBO 来近似估计扩散模型的似然(Likelihood)。
Related Works
扩散模型作为研究热点,现阶段已有部分工作探索在其上的成员推理,然而并不能良好适配文生图扩散模型:
[5]
提出了基于似然比(Likelihood Ratio Attack)的成员推理方法,然而该方法由于需要训练大量阴影模型(Shadow Model)导致存在高计算开销,无法扩展(Scale-up)到文生图扩散模型上。
[6,7,8]
提出了基于查询的成员推理,计算开销更小,可以扩展到文生图扩散模型。但是由于评估设定不合理而导致的成功幻觉,使其在更真实的文生图任务场景下达不到相对满意的效果。
本文方法
针对现有挑战,本文提出了一种基于
条件似然差异
(
C
onditional
Li
kelihood
D
iscrepancy, CLiD)的成员推理方法。
[1] BBC. 'Art is dead Dude' - the rise of the AI artists stirs debate. 2022. URL https://www.bbc.com/news/technology-62788725.
[2] CNN. AI won an art contest, and artists are furious. 2022. URL https://www.cnn.com/2022/09/03/ tech/ai-art-fair-winner-controversy/index.html.
[3] Reuters. Lawsuits accuse AI content creators of misusing copyrighted work. 2023. URL https://www.reuters.com/legal/transactional/ lawsuits-accuse-ai-content-creators-misusing-copyrighted-work-2023-01-17/.
[4] WashingtonPost. He made a children’s book using AI. Then came the rage. 2022. URL https://www.washingtonpost.com/technology/2023/01/19/ ai-childrens-book-controversy-chatgpt-midjourney/.
[5] Nicolas Carlini et al. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23)
[6] Jinhao Duan et al. Are diffusion models vulnerable to membership inference attacks? In International Conference on Machine Learning, 2023.
[7] Fei Kong et al. An efficient membership inference attack for the diffusion model by proximal initialization. In The Twelfth International Conference on Learning Representations, 2024.
[8] Wenjie Fu et al. A probabilistic fluctuation based membership inference attack for generative models. arXiv preprint arXiv:2308.12143, 2023
[9] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. 'Denoising diffusion probabilistic models.' Advances in neural information processing systems 33 (2020): 6840-6851.
[10] Rombach, Robin, et al. 'High-resolution image synthesis with latent diffusion models.' Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.
[11] Li, Alexander C., et al. 'Your diffusion model is secretly a zero-shot classifier.' Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.
[12] Debeshee Das, Jie Zhang, and Florian Tramèr. Blind baselines beat membership inference attacks for foundation models. arXiv preprint arXiv:2406.16201, 2024.
[13] Dubiński, Jan, et al. 'Towards more realistic membership inference attacks on large diffusion models.' Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024.
[14] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3–18. IEEE, 2017.
[15] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897–1914. IEEE, 2022.