De-amplifying Bias from Differential Privacy in Language Model Fine-tuning

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Srivastava, Sanjari, Mardziel, Piotr, Zhang, Zhikhun, Ahlawat, Archana, Datta, Anupam, Mitchell, John C
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866909096398028800
author Srivastava, Sanjari
Mardziel, Piotr
Zhang, Zhikhun
Ahlawat, Archana
Datta, Anupam
Mitchell, John C
author_facet Srivastava, Sanjari
Mardziel, Piotr
Zhang, Zhikhun
Ahlawat, Archana
Datta, Anupam
Mitchell, John C
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2402_04489
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle De-amplifying Bias from Differential Privacy in Language Model Fine-tuning
Srivastava, Sanjari
Mardziel, Piotr
Zhang, Zhikhun
Ahlawat, Archana
Datta, Anupam
Mitchell, John C
Machine Learning
Cryptography and Security
Computers and Society
Methodology
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.
title De-amplifying Bias from Differential Privacy in Language Model Fine-tuning
topic Machine Learning
Cryptography and Security
Computers and Society
Methodology
url https://arxiv.org/abs/2402.04489