Federated Learning Model Training with Label Bias Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning models face limitations due to biased training data sets, leading to reduced accuracy and efficiency in model training, as some labels are underrepresented or overrepresented, affecting the robustness and performance of global models.
Innovation Solution
A system identifies computing nodes with biased local training data by analyzing statistical data and provides additional training data to balance the representation of labels, using data augmentation techniques to mitigate bias without requiring access to sensitive local data, thereby improving the accuracy and speed of model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train models across decentralized computing nodes without data sharing, then data privacy and security are improved, but model training accuracy deteriorates due to biased local training data sets
Solution Approach 1:
A central server acts as an intermediary to collect statistical data from decentralized computing nodes, analyze label distribution bias, and redistribute balanced training data samples. This mediator coordinates between nodes without requiring direct data sharing, maintaining privacy while improving accuracy through centralized statistical analysis and data balancing.
Solution Approach 2:
The system changes the parameter of label distribution across training data sets by identifying biased nodes and redistributing data samples to achieve more uniform label representation. This parameter adjustment transforms imbalanced local data sets into balanced ones, improving model training accuracy while preserving the decentralized federated learning architecture.
2Measurement precision
If additional training data is provided to computing nodes with biased data sets, then model accuracy is improved, but computational resources and training time increase
Solution Approach 1:
Instead of redistributing all training data, the system applies partial action by selectively providing only the specific data samples needed to balance label representation at each computing node. This targeted approach avoids the excessive computational overhead of complete data redistribution while achieving sufficient balance for accurate model training.
Solution Approach 2:
The system performs preliminary analysis of statistical data and label distribution bias before data redistribution. By pre-identifying which nodes have biased data sets and what specific labels need balancing, the system prepares an optimized data allocation plan that minimizes unnecessary data transfer and computational resources during the actual training process.
3Measurement precision
If statistical data analysis is performed to identify biased computing nodes, then data distribution quality is improved, but system complexity increases
Solution Approach 1:
The system extracts only the essential statistical data (label counts and distributions) from each computing node's local training data, rather than analyzing the complete data sets. This extraction approach simplifies the analysis process by focusing on key metrics that indicate bias, reducing computational complexity while maintaining effective identification of problematic nodes.
Data Source
AI summary
A computer system trains a federated learning model. A federated learning model is distributed to a plurality of computing nodes, each having a set of local training data comprising labeled data samples. Statistical data is received from each computing node that indicates the node's count of data samples for each label, and is analyzed to identify one or more computing nodes having local training data in which a label category is underrepresented beyond a threshold value with respect to data samples. Additional data samples labeled with the underrepresented labels are provided, and the computing nodes perform training. Results of training are received and are processed to generate a trained global model. Embodiments of the present invention further include a method and program product for training a federated learning model in substantially the same manner described above.