BlackMirror: Black-Box Backdoor Detection for Text-to-Image Models via Instruction-Response Deviation

Researchers have developed BlackMirror, a new framework for detecting backdoored text-to-image AI models in black-box settings. The system identifies semantic deviations between visual patterns and instructions, offering a training-free solution that can be deployed in Model-as-a-Service applications.

• →BlackMirror introduces a novel approach to detect backdoor attacks in text-to-image models without requiring model access or training.
• →The framework identifies partial semantic manipulations in generated images while distinguishing them from benign content variations.
• →MirrorMatch component aligns visual patterns with instructions to detect semantic deviations from expected outputs.
• →MirrorVerify evaluates stability of deviations across different prompts to confirm true backdoor behavior.
• →The system addresses limitations of existing detection methods that rely on image-level similarity and struggle with diverse backdoor attacks.