De-amplifying Bias from Differential Privacy in Language Model Fine-tuning
This addresses the challenge for ML practitioners in balancing privacy and fairness in trustworthy AI, though it is incremental as it applies an existing method (CDA) to a new problem.
The study tackled the problem that differential privacy (DP) amplifies gender, racial, and religious bias in fine-tuned large language models, making them more biased than non-DP models, and showed that Counterfactual Data Augmentation (CDA) mitigates this amplification, enabling models to maintain both fairness and privacy.
Fairness and privacy are two important values machine learning (ML) practitioners often seek to operationalize in models. Fairness aims to reduce model bias for social/demographic sub-groups. Privacy via differential privacy (DP) mechanisms, on the other hand, limits the impact of any individual's training data on the resulting model. The trade-offs between privacy and fairness goals of trustworthy ML pose a challenge to those wishing to address both. We show that DP amplifies gender, racial, and religious bias when fine-tuning large language models (LLMs), producing models more biased than ones fine-tuned without DP. We find the cause of the amplification to be a disparity in convergence of gradients across sub-groups. Through the case of binary gender bias, we demonstrate that Counterfactual Data Augmentation (CDA), a known method for addressing bias, also mitigates bias amplification by DP. As a consequence, DP and CDA together can be used to fine-tune models while maintaining both fairness and privacy.