Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• BlackMirror is a novel framework for detecting backdoored text-to-image models in black-box settings.method 95%
• Existing backdoor detection approaches for text-to-image models struggle to generalize to backdoor attacks with visually diverse generations.limitation 90%
• BlackMirror is motivated by the observation that only partial semantic patterns within generated images are steadily manipulated across backdoor attacks, while the rest of the content remains diverse or benign.technical 95%
• BlackMirror consists of two components: MirrorMatch and MirrorVerify.method 95%
• MirrorMatch aligns visual patterns with corresponding instructions to detect semantic deviations.method 95%
• MirrorVerify evaluates the stability of deviations across varied prompts to distinguish true backdoor behavior from benign responses.method 95%
• BlackMirror is a general, training-free framework deployable as a plug-and-play module in Model-as-a-Service (MaaS) applications.market 90%
• BlackMirror achieves accurate detection across a wide range of attacks.result 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card