Abstract

With language models becoming increasingly embedded in our digital interactions, concerns over data privacy and information leakage are becoming more apparent. The lack of regulation and transparency in how pipelines involving LLMs collect and use data can exacerbate these privacy risks. This talk provides a brief overview of privacy issues in language models and interventions (such as differential privacy) that have been proposed to mitigate privacy leakage. Finally, we highlight challenges in privacy-preserving research and outline future directions that may help drive advancements in this field.