Group Bias Mitigation in Federated Learning via Fairness Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face challenges in detecting and mitigating group bias in model training data, particularly when data is sensitive and not centrally accessible, leading to potential accuracy issues when dealing with underrepresented datasets.
Innovation Solution
A supervisory device in a federated learning system computes accuracy and fairness loss metrics for an aggregated model, initiating additional training rounds to optimize a tradeoff between accuracy and fairness through a constrained optimization problem, without requiring access to the underlying training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train models at distributed sites with local data, then data privacy is protected and data sharing is avoided, but model fairness becomes challenging and group bias cannot be effectively detected or mitigated
Solution Approach 1:
The patent introduces intermediary components (fairness evaluator and bias mitigation module) that act as mediators between the distributed training sites and the global model aggregation. These intermediaries receive model parameters from various sites, evaluate fairness metrics without accessing raw data, and generate correction signals that are fed back to adjust training at individual sites, thus resolving the contradiction between privacy protection and bias detection
Solution Approach 2:
The system implements a feedback mechanism where fairness metrics are continuously evaluated on the aggregated model and the results are fed back to the distributed training sites. This feedback loop allows the system to detect and mitigate group bias while maintaining data privacy, as the feedback provides guidance for improving fairness without requiring centralization of sensitive data
2Reliability
If model training is performed at distributed sites without centralizing data, then data privacy is maintained, but fairness-related metrics become difficult to compute and optimize
Solution Approach 1:
The fairness evaluator serves as an intermediary that computes fairness metrics by receiving only aggregated model parameters from the supervisory device, not the raw training data. This intermediary computes fairness-related metrics (such as demographic parity or equalized odds) based on the model's predictions on representative datasets, enabling precise fairness measurement while preserving data privacy at distributed sites
Solution Approach 2:
The system uses representative datasets or synthetic data copies that mirror the demographic composition of the overall population to compute fairness metrics. Instead of requiring access to actual sensitive training data, the system creates copies or approximations that preserve the statistical properties needed for fairness evaluation, thus enabling accurate fairness measurement without compromising privacy
3Reliability
If additional training rounds are initiated to improve fairness, then group bias mitigation is enhanced, but training time and computational resources increase
Solution Approach 1:
The system applies partial retraining rather than complete retraining in additional training rounds. The bias mitigation module identifies specific model parameters or feature representations that contribute to bias and applies targeted adjustments, rather than retraining the entire model from scratch. This partial action approach reduces the time and computational resources required while still effectively mitigating group bias
Solution Approach 2:
The fairness evaluator continuously monitors fairness metrics during the training process and triggers additional training rounds only when fairness thresholds are violated or when improvement is detected. This preliminary evaluation and conditional triggering prevents unnecessary additional training rounds, optimizing the balance between fairness improvement and training time consumption
Data Source
AI summary
In one embodiment, a supervisory device in a federated learning system generates an aggregated model that aggregates a plurality of machine learning models trained by trainer nodes in a federated learning system during a training round. The supervisory device computes an accuracy loss metric for the aggregated model. The supervisory device also computes a fairness loss metric for the aggregated model based on fairness-related metrics associated with the plurality of machine learning models trained by the trainer nodes. The supervisory device initiates an additional training round during which the trainer nodes retrain their machine learning models for aggregation by the apparatus, in accordance with a constrained optimization problem that seeks to optimize a tradeoff between accuracy and fairness associated with the aggregated model.