Robust Backdoor Attack with Visible, Semantic, Sample-specific and Compatible Triggers

Ruotong Wang, Hongrui Chen, Zihao Zhu, Li Liu, Yong Zhang, Yanbo Fan, Baoyuan Wu

Primary Area: societal considerations including fairness, safety, privacy

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.

Keywords: backdoor attack

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Deep neural networks (DNNs) can be manipulated to exhibit specific behaviors when exposed to specific trigger patterns, without affecting their performance on benign samples, dubbed backdoor attack.
Some recent research has focused on designing invisible triggers for backdoor attacks to ensure visual stealthiness, while showing high effectiveness, even under backdoor defense.
However, we find that these carefully designed invisible triggers are often susceptible to visual distortion during inference, such as Gaussian blurring or environmental variations in physical scenarios. This phenomenon could significantly undermine the practical effectiveness of attacks, but has been rarely paid attention and thoroughly investigated.
To address this limitation, we define a novel trigger called the Visible, Semantic, Sample-Specific, and Compatible trigger (VSSC-trigger), to achieve effective, stealthy and robust simultaneously.
To implement it, we develop an innovative approach by utilizing the powerful capabilities of large language models for choosing the suitable trigger and advanced image editing technique for generating the poisoned image with the trigger.
Extensive experimental results and analysis validate the effectiveness, stealthiness and robustness of the VSSC-trigger. It demonstrates superior resistance to distortions compared to most digital backdoor attacks and allows more efficient and flexible trigger integration compared to physical backdoor attacks.
We hope that the proposed VSSC trigger and implementation approach could inspire future studies on designing more practical triggers in backdoor attack.

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: 5081