DiSCO: A Black-Box Defense Against Harmful Text-to-Image Generation
A new defense mechanism called DiSCO has been developed by researchers to protect text-to-image generative models from risks associated with Not-Safe-For-Work (NSFW) content, such as violence and nudity, especially in light of red-teaming adversarial attacks. DiSCO functions solely at the prompt level as a plug-and-play solution, eliminating the need for model retraining, fine-tuning, or internal component access, making it adaptable for proprietary models. In contrast to existing white-box defenses that depend on text encoder adjustments or inference-time modifications, current black-box strategies utilizing LLM prompt rewriting struggle with the "benign adversarial" issue, where prompts appear safe yet produce harmful outcomes. DiSCO counters this with distribution-guided contrastive prompt optimization. The study can be found on arXiv with the identifier 2608.17067.
Key facts
- DiSCO is a zero-shot, strictly black-box defense for text-to-image generation.
- It works at the prompt level without model retraining, fine-tuning, or access to model internals.
- Targets NSFW content such as violence and nudity generated by text-to-image models.
- Existing white-box defenses rely on text encoder optimization, weight editing, or inference-time intervention.
- Black-box LLM-based prompt rewriting fails against the 'benign adversarial' problem.
- Benign adversarial prompts are linguistically safe but still trigger harmful generation due to learned data distributions.
- DiSCO uses distribution-guided contrastive prompt optimization.
- Paper published on arXiv with identifier 2608.17067.
Entities
—