VinAIResearch/Anti-DreamBooth

Implements multiple perturbation algorithms (ASPL, FSMG, and ensemble variants) that inject imperceptible noise into images to degrade DreamBooth fine-tuning quality while maintaining visual fidelity. Built on Hugging Face diffusers and compatible with multiple Stable Diffusion versions (1.4–2.1), with recent LoRA support for parameter-efficient personalization attacks. Evaluated on facial datasets (VGGFace2, CelebA-HQ) with robustness testing across model/prompt mismatches.

267 stars. No commits in the last 6 months.