Exploring the Unfairness of DP-SGD Across Settings
This work addresses the problem of balancing privacy and fairness for end users and regulators, but it is incremental as it builds on prior suggestions of conflicts between these objectives.
The study investigated the trade-off between privacy and fairness in AI models by evaluating DP-SGD across different settings, finding a negative logarithmic correlation for linear classification and robust deep learning, with no significant impact on fairness for PCA.
End users and regulators require private and fair artificial intelligence models, but previous work suggests these objectives may be at odds. We use the CivilComments to evaluate the impact of applying the {\em de facto} standard approach to privacy, DP-SGD, across several fairness metrics. We evaluate three implementations of DP-SGD: for dimensionality reduction (PCA), linear classification (logistic regression), and robust deep learning (Group-DRO). We establish a negative, logarithmic correlation between privacy and fairness in the case of linear classification and robust deep learning. DP-SGD had no significant impact on fairness for PCA, but upon inspection, also did not seem to lead to private representations.