Privacy-Preserving Federated Learning via Masked Model Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face challenges in maintaining data privacy, as they can inadvertently expose personally identifiable information or sensitive data, even when client models are not explicitly shared, due to indirect inferences from combined model data, which violates regulatory requirements and diminishes model accuracy through anonymization or synthetic data approaches.
Innovation Solution
The implementation of a privacy-preserving federated learning method that uses masked client models, where data values are cryptographically obfuscated, and only unmasked with a secure unmasking data object, ensuring that intermediary systems cannot access true values without cooperation from the mask combiner system and the privacy-preserving federated learning system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If federated learning systems aggregate distributed models to build a global model, then model accuracy and learning effectiveness are improved, but privacy leakage risk increases due to indirect inferences from combined model data
Solution Approach 1:
The patent introduces masked models as an intermediary representation between client local models and the global model aggregation process. These masked models contain obfuscated parameter values that prevent direct inference of sensitive information while preserving the mathematical properties needed for federated averaging, thus resolving the contradiction between model accuracy and privacy protection
Solution Approach 2:
The patent transforms model parameters by applying cryptographic masking operations, changing the state of parameters from their original form to a masked form that cannot be directly reverse-engineered. This parameter transformation maintains the utility of parameters for model aggregation while eliminating privacy vulnerabilities
2Object-affected harmful factors
If client models are anonymized or synthetic data is used to protect privacy, then data privacy is improved, but model accuracy deteriorates due to loss of information
Solution Approach 1:
The patent creates masked copies of client model parameters that serve as privacy-preserving substitutes for the original parameters. These copies maintain the statistical and mathematical properties necessary for accurate model training and aggregation, while being computationally infeasible to reverse to the original sensitive data
3Object-affected harmful factors
If masked models are used to preserve privacy, then privacy protection is improved, but system complexity increases due to additional masking and unmasking operations
Solution Approach 1:
The patent applies masking operations to client models before they are transmitted to the server for aggregation, performing the privacy protection action in advance. This preliminary masking eliminates the need for complex post-processing and security infrastructure at the server side, as the privacy protection is already embedded in the transmitted parameters
Data Source
AI summary
Privacy-preserving federated learning apparatuses, systems, computer program products, and methods are provided that generate an updated global model based on a set of client models while maintaining privacy regarding the data values embodying each client model and the updated global model. In this regard, masked client models are utilized, which cryptographically obfuscate data values embodying the client model while still enabling combination, or “aggregation,” of the masked client models to generate a masked updated global model. The masked updated global model similarly includes obfuscated data values embodying the updated global model, but may be unmasked to reveal the true values of the updated global model for use. Some embodiments utilize specific steps for communication between environments, systems, devices, and/or the like, to ensure the masked models can only be unmasked by intended entities.