Awakening the Hydra: Stabilizing Multi-Concept Backdoor Injection in Text-to-Image Diffusion Models
Researchers have developed Hydra, a framework designed to stabilize multi-concept backdoor injections in text-to-image diffusion models. This is crucial because open-source models are often fine-tuned and redistributed, leading to potential conflicts and degraded quality from accumulated backdoor behaviors. Hydra addresses this by evolving text encoder triggers that align with target concepts while remaining stable across others, and uses multi-task fine-tuning with regularization to enhance training stability. Experiments show Hydra achieves high attack success rates while preserving clean generation fidelity. AI
IMPACT Introduces a method to control and stabilize backdoor injections in diffusion models, impacting model security and trustworthiness.
