Detecting CSAM Text-to-Image LoRAs From Weights

2026-07-28

AI Summary

Current methods for auditing LoRA models trained to generate child sexual abuse material (CSAM) rely on metadata or generated content, rendering them susceptible to evasion and entailing legal and ethical risks. This work proposes a novel inference-free, weight-level detection approach that extracts the dominant left singular vector (u₁) from LoRA weights via singular value decomposition as a fingerprint to directly identify the model’s training objective. To our knowledge, this is the first method capable of extracting a secure and robust fingerprint signal solely from model weights—without generating images or depending on metadata. In experiments using human age as a proxy task, the method accurately identifies harmful training objectives, generalizes effectively across base models, automatically abstains on irrelevant benign content, and demonstrates strong robustness against additive noise, scaling, and reduced numerical precision.

Abstract

Low-rank adaptation (LoRA) fine-tuning has made it cheap and easy to customize open-weight image generation models for specific tasks, including the production of child sexual abuse material (CSAM). Existing moderation relies on metadata or generated outputs, but metadata can be deceptive and generating outputs may itself be unacceptable or illegal. We show that a safer signal lives in the weights. The top-left singular vectors of a LoRA's updates form a compact, inference-free fingerprint ($u_1$) of its strongest learned change. Using human-subject age as a benign proxy for CSAM, we find that $u_1$ 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. These results indicate that harmful LoRAs could be screened directly from their weights without relying on metadata or generating harmful outputs.

Problem

• CSAM
• LoRA
• weight-based detection
• text-to-image generation
• model moderation

Innovation

• LoRA
• weight-based detection
• singular vectors
• CSAM detection
• model moderation