How Does Differential Privacy Affect Social Bias in LLMs? A Systematic Evaluation

Systematic evaluation of social bias in LLMs trained with DP-SGD across four complementary bias metrics, investigating the relationship between differential privacy and fairness in large language models.