Reproduce “Semantic-level Backdoor Attack against Text-to-Image Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianxin Chen, Wenbo Jiang, Hongqiao Chen, Zhirun Zheng, Cheng Huang
A new semantic-level backdoor attack, SemBD, targets Text-to-Image diffusion models by operating in the continuous semantic representation space, rather than discrete textual patterns. This approach enables the attack to achieve high success rates and evade state-of-the-art input-level and attention-consistency-based defenses while preserving model utility.
Key insights
- Backdoor triggers defined in the continuous semantic representation space offer significantly higher stealthiness and robustness against current input-level defenses compared to discrete textual triggers.
- Multi-entity backdoor targets are crucial for diversifying cross-attention patterns, effectively circumventing attention-consistency-based detection mechanisms.
- Semantic regularization is essential for ensuring trigger specificity, preventing unintended activation by prompts that contain only partial semantic information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.