18 hours ago
- LoRA fine-tuning enables cheap customization of open-weight image generation models for CSAM production.
- Existing moderation methods rely on metadata or generated outputs, but metadata is deceptive and output generation may be unacceptable or illegal.
- A safer detection signal exists in the weights: the top-left singular vectors of LoRA updates form a compact, inference-free fingerprint (u1) of the strongest learned change.
- The u1 fingerprint identifies what a LoRA was trained on, generalizes across base models, and abstains on unrelated benign content.
- The signal is robust to additive weight noise, rescaling, and precision reduction.
- Harmful LoRAs can be screened directly from weights without metadata or generating harmful outputs.