ML Bias Assessment Using Training Metadata Ledger
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often exhibit bias due to inadequate training data, which is difficult to identify post-deployment due to data privacy and sovereignty concerns, leading to poor performance and ethical issues.
Innovation Solution
A device retrieves metadata from a ledger associated with the machine learning model to assess bias by comparing input data to training data metadata, providing an indication of bias for display and potentially controlling inference decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is kept private to protect data sovereignty and privacy, then data security and compliance are improved, but the ability to identify and assess model bias deteriorates
Solution Approach 1:
The patent extracts only the necessary metadata about training data (such as data characteristics, distribution statistics, and composition information) while leaving the actual training data private and inaccessible. This allows bias assessment without compromising data security or privacy, resolving the contradiction between data protection and bias detection capability.
Solution Approach 2:
The patent introduces metadata as an intermediary between the private training data and the bias assessment process. This metadata serves as a mediator that contains sufficient information to evaluate model bias while maintaining the privacy and security of the original training data, enabling bias detection without data exposure.
2Object-affected harmful factors
If training data is concealed due to privacy laws, then data privacy compliance is improved, but model performance monitoring deteriorates
Solution Approach 1:
The patent extracts only the essential metadata needed for performance monitoring and bias assessment, separating this information from the confidential training data. This allows continuous monitoring of model performance and detection of bias while maintaining full privacy compliance, as the actual training data remains concealed.
Solution Approach 2:
The patent performs preliminary assessment of training data characteristics through metadata capture during the training process. This preliminary action enables future performance monitoring and bias detection without requiring access to the training data at later stages, maintaining privacy compliance while ensuring model reliability.
3Difficulty of detecting and measuring
If metadata about training data is made accessible, then bias assessment capability is improved, but data sovereignty and privacy protection deteriorate
Solution Approach 1:
The patent applies local quality by providing metadata with specific, targeted characteristics that are sufficient for bias assessment but do not reveal sensitive information about the training data. The metadata includes localized information such as data distribution statistics and composition details without exposing the actual data content, maintaining privacy while enabling assessment.
Solution Approach 2:
The patent extracts only the minimal necessary metadata for bias assessment purposes, separating this information from the confidential training data. This extraction approach provides the needed assessment capability while preserving privacy protection, as the extracted metadata contains no sensitive information about the actual training data.
Data Source
AI summary
In one embodiment, a device receives a request for a machine learning model to make an inference about input data included in the request. The device retrieves metadata regarding training data used to train the machine learning model from a ledger associated with the machine learning model. The device assesses bias of the machine learning model by comparing the input data in the request to the metadata from the ledger. The device provides an indication of the bias of the machine learning model for display.