How Does Differential Privacy Affect Social Bias in LLMs? A Systematic Evaluation
This work addresses the underexplored relationship between DP and social bias in LLMs, providing a multi-paradigm evaluation framework for practitioners concerned with both privacy and fairness.
The study evaluates how differential privacy (DP) affects social bias in LLMs by comparing a DP-trained model with non-DP baselines across four tasks. It finds that DP reduces bias in sentence scoring but not consistently across tasks, and that reduced memorization does not guarantee reduced unfairness.
Large language models (LLMs) trained on web-scale corpora can memorize sensitive training data, posing significant privacy risks. Differential privacy (DP) has emerged as a principled framework that limits the influence of individual data points during training, yet the relationship between differential privacy and social bias in LLMs remains poorly understood. To investigate this, we present a systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing a DP model against non-DP baselines across four complementary paradigms: sentence scoring, text completion, tabular classification, and question answering. We find that DP reduces bias in sentence scoring tasks, where bias is measured through controlled likelihood comparisons, yet this improvement does not generalize across all tasks. Our results reveal a discrepancy between logit-level bias and output-level bias. Moreover, decreasing memorization does not necessarily reduce unfairness, underscoring the importance of multi-paradigm evaluation when assessing fairness in LLMs.