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 protection for sensitive training data. PrivBoN uses Gumbel noise to achieve differential privacy and KL-regularized alignment, matching theoretical performance under certain privacy budgets. PrivITP further refines this by combining $\chi^2$-regularized rejection sampling with a Gaussian mechanism, offering ex-post $(\epsilon,\delta)$-DP independent of response count and decoupling regularization from privacy parameters. AI
IMPACT These methods could enable more secure deployment of LLMs by protecting sensitive user data during inference.
RANK_REASON Academic paper detailing new methods for differential privacy in LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
