AI Post-Processing for Bias Mitigation Without Sensitive Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI/ML models often produce biased predictions due to biased training datasets, which can lead to unfair outcomes, and conventional bias mitigation techniques require access to sensitive attributes or modify the model architecture, posing privacy and regulatory challenges.
Innovation Solution
A post-processing method that segregates data samples based on predicted probability scores and loss values to identify and transition labels of low-confidence and high-loss samples, mitigating bias without accessing sensitive attributes or modifying the model architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional post-processing techniques are used to mitigate bias, then bias reduction is achieved, but access to sensitive attributes is required which violates privacy and regulatory constraints
Solution Approach 1:
The patent extracts and removes sensitive attribute information from the processing pipeline entirely. Instead of using sensitive attributes for bias mitigation, the system identifies and corrects biased predictions based solely on non-sensitive features and prediction confidence metrics, thereby eliminating the privacy violation while maintaining bias reduction capability
Solution Approach 2:
The patent introduces an intermediary mechanism that acts as a mediator between the model predictions and the final decisions. This intermediary component analyzes prediction confidence scores and loss values to identify potentially biased predictions, then applies correction rules without requiring access to sensitive attributes, thus resolving the contradiction between bias reduction and privacy protection
2Reliability
If in-processing techniques are used to mitigate bias, then fair decisions are achieved, but model architecture modification is required which increases complexity and training time
Solution Approach 1:
The patent segments the bias mitigation process into a separate post-processing stage rather than integrating it into the model architecture. This segmentation allows the base model to remain simple and unchanged, while a separate correction module handles fairness considerations by analyzing prediction confidence and loss values, thus eliminating the need for architectural modification
Solution Approach 2:
The patent performs preliminary analysis of prediction confidence scores and loss values before final predictions are made. By pre-identifying predictions that exhibit bias patterns through confidence and loss thresholding, the system prepares correction data in advance, allowing fair decisions to be achieved without modifying the underlying model architecture or training process
3Reliability
If pre-processing techniques are used to balance training data distribution, then biased predictions are prevented, but sensitive attribute information must be processed which creates privacy risks
Solution Approach 1:
The patent extracts and eliminates sensitive attribute processing from the pre-processing stage. Instead of balancing data distribution using sensitive attributes, the system uses non-sensitive features and prediction metrics to identify and correct biased outcomes, thereby preventing bias while removing the privacy risk associated with handling sensitive information
Solution Approach 2:
The patent employs temporary, disposable correction rules applied only during the post-processing stage rather than maintaining sensitive attribute information throughout the entire pipeline. The correction mechanism uses transient confidence scores and loss values that are discarded after application, eliminating the need to retain or process sensitive attributes while achieving prediction fairness
Data Source
AI summary
Methods and systems for mitigating bias from Artificial Intelligence (AI) models during post-processing are disclosed. Method performed by a server system includes accessing first predicted probability score and second predicted probability score for each data sample which is associated with a particular predicted class label having a counterpart class label, a predicted loss category label. Method includes segregating a first set of data samples based on the first predicted probability score and a predefined margin threshold. Method includes segregating a second set of data samples based on the second predicted probability score and a first predefined loss threshold. Method includes segregating a third set of data samples based on the first set of data samples, the second set of data samples, and an overlap condition. Method includes transitioning the predicted class label of each of the third set of data samples to the counterpart class label.