Selective Fine-Tuning for Targeted and Robust Concept Unlearning
TRUST is a method for selectively removing harmful concepts from text-to-image models using fine-tuning. It dynamically identifies concept neurons within the model and uses two regularization objectives—one for hard unlearning and one for soft unlearning—guided by Hessian regularization. This approach successfully removes target concepts while preserving the quality of benign image generation and improving robustness against adversarial prompts.