DP Model Training for Fair Group Privacy Cost and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing differential-privacy training approaches fail to effectively measure and mitigate group-group disparities in privacy costs and model accuracy, leading to higher privacy costs for underrepresented data types and inaccurate model training.
Innovation Solution
A modified differential-privacy training process that adjusts per-sample gradients adaptively, using clipping and scaling based on a reference bound, and adds noise, while evaluating gradient direction changes to balance privacy and accuracy across groups without requiring group labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If per-sample gradients are clipped and noise is added in differential-privacy training, then individual data sample privacy is protected, but group-group disparities in privacy cost and model accuracy increase
Solution Approach 1:
The patent applies local quality by differentiating gradient clipping and noise addition based on group characteristics. Instead of uniform differential privacy application, the system identifies underrepresented groups and applies adjusted clipping bounds and noise scales specific to each group's gradient direction errors, thereby protecting privacy while maintaining group-specific model accuracy
Solution Approach 2:
The system dynamically adjusts differential privacy parameters (clipping bounds C0, noise scale σ) based on measured gradient direction errors for different groups. By changing these parameters adaptively according to group performance metrics, the system resolves the contradiction between privacy protection and model accuracy for underrepresented groups
2Loss of information
If existing differential-privacy approaches are used, then privacy cost is bounded for individual samples, but group-group differences in privacy cost cannot be measured or corrected
Solution Approach 1:
The patent segments the privacy cost measurement into group-specific components by calculating gradient direction errors separately for each group label. This segmentation enables the system to identify and measure disparities in privacy costs across different groups, revealing inequalities that uniform measurement would miss
Solution Approach 2:
The system implements feedback by measuring gradient direction errors for each group and using these measurements to adjust differential privacy parameters. The measured group privacy costs feed back into the training process, enabling continuous correction of group-group disparities in privacy protection
3Reliability
If uniform clipping bound is applied to all samples, then differential privacy guarantee is maintained, but underrepresented groups suffer higher privacy costs and gradient direction errors
Solution Approach 1:
The patent transforms the static uniform clipping bound into a dynamic group-specific clipping bound. The system calculates separate clipping bounds for different groups based on their gradient direction errors, allowing the clipping mechanism to adapt to each group's characteristics while maintaining overall differential privacy guarantees
Solution Approach 2:
By applying different clipping bounds to different groups based on their specific gradient direction errors, the system implements local quality. Underrepresented groups receive adjusted clipping parameters that preserve their gradient direction accuracy, while represented groups maintain their privacy protection
Data Source
AI summary
A model evaluation system evaluates the extent to which privacy-aware training processes affect the direction of training gradients for groups. A modified differential-privacy (“DP”) training process provides per-sample gradient adjustments with parameters that may be adaptively modified for different data batches. Per-sample gradients are modified with respect to a reference bound and a clipping bound. A scaling factor may be determined for each per-sample gradient based on the higher of the reference bound or a magnitude of the per-sample gradient. Per-sample gradients may then be adjusted based on a ratio of the clipping bound to the scaling factor. A relative privacy cost between groups may be determined as excess training risk based on a difference in group gradient direction relative to an unadjusted batch gradient and the adjusted batch gradient according to the privacy-aware training.