LORA (Low-Rank Adaptation) is a technique for fine-tuning large AI models on small datasets. In the context of NCII, perpetrators use LORA to fine-tune Stable Diffusion or similar models on 10-30 photos of a specific individual, creating a custom model capable of generating realistic intimate imagery of that person. LORA models targeting specific individuals are shared on model-sharing platforms (CivitAI, HuggingFace) where other users can download them for free. The creation and distribution of intimate content using such a model are concerning.

Key facts about this term

  1. Understand the threat model Just 10-30 publicly available photos of you can be used to train a personalized LORA model.
  2. Report LORA models targeting you to hosting platforms CivitAI and HuggingFace have content policies against LORA models created for NCII purposes.
  3. Use ScanErase to identify LORA-generated content The biometric scan identifies LORA-generated deepfake content using facial geometry matching.

Frequently asked questions

Can platforms hosting LORA models be required to remove models targeting me?

Model-sharing platforms have content policies covering NCII-purpose models.

Can I reduce the risk of LORA models being created from my photos?

Reducing publicly available high-quality face photographs reduces the training material available. Photo cloaking tools (Fawkes, Glaze) add adversarial noise that disrupts AI training in some cases.

Related glossary terms