Deploy Securely
Sensitive Data Generation
6:40
I’m worried about data leakage from LLMs, but probably not why you think. While unintended training is a real risk that can’t be ignored, something else is going to be a much more serious problem: sensitive data generation (SDG). A recent paper (https://arxiv.org/pdf/2310.07298v1.pdf) shows how LLMs can infer huge amounts of personal information from seemingly innocuous comments on Reddit. And…