Reproduce “Low-Effort Jailbreak Attacks Against Text-to-Image Safety Filters” 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
Ahmed B Mustafa, Zihan Ye, Yang Lu, Michael P Pound, Shreyank N Gowda
This research from the University of Nottingham and collaborators systematically categorizes and evaluates low-effort, natural language-based jailbreak attacks against text-to-image models. The study reveals that these simple prompt engineering strategies consistently bypass state-of-the-art safety filters, with some achieving attack success rates above 70%, highlighting widespread vulnerabilities in current T2I moderation pipelines.
Key insights
- Current T2I safety filters are broadly vulnerable to accessible, natural language-based prompt manipulations, exploiting semantic ambiguities and contextual reframing.
- The 'Material Substitution Attack', involving benign material descriptors, was particularly effective at bypassing safety mechanisms while still generating restricted content.
- Vulnerabilities to these low-effort attacks are systemic across diverse T2I models, indicating a broader challenge in aligning generative models with safety policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.