PrivITP

New methods enhance LLM privacy and alignment without performance loss

Researchers have developed new methods, PrivBoN and PrivITP, to enhance differential privacy in large language models during inference. These techniques address issues like reward hacking and the lack of privacy protect…