Data Exfiltration in Diffusion Models: A Backdoor Attack Approach

Cheng Yu Yeo, Yi-Lun Wu, Hong-Han Shuai

Keywords: Backdoor Attack, Data Exfiltration, Diffusion Model

Abstract: As diffusion models (DMs) become increasingly susceptible to adversarial attacks, this paper investigates a novel method of data exfiltration through strategically implanted backdoors. Unlike conventional techniques that directly alter data, we pioneer the use of unique trigger embeddings for each image to enable covert data retrieval. Furthermore, we extend our exploration to text-to-image diffusion models such as Stable Diffusion by introducing the Caption Backdoor Subnet (CBS), which exploits these models for both image and caption extraction. This innovative approach not only reveals an unexplored facet of diffusion model security but also contributes valuable insights toward enhancing the resilience of generative models against sophisticated threats.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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.

Submission Number: 3806