Encrypted Federated Model Aggregation for Private ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training machine learning models using sensitive data from multiple sources face challenges such as the potential leakage of personal information, computational overhead, and complexity in auditing, especially in federated learning and homomorphic encryption approaches.
Innovation Solution
A method combining federated learning with homomorphic encryption, where local trained functions are encrypted and aggregated to form a global encrypted dataset, allowing for secure transmission and calculation without revealing personal information, using homomorphic encryption algorithms to maintain data privacy and reduce computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is used for complete training on encrypted data, then data privacy is improved, but processing costs and algorithm limitations worsen
Solution Approach 1:
The training process is divided into two distinct phases: (1) local training phase where each client system trains local models on their own data, and (2) aggregation phase where encrypted local parameters are combined server-side. This segmentation allows encryption to be applied only where necessary (during aggregation and storage) rather than throughout the entire training process, significantly reducing computational overhead while maintaining privacy guarantees.
Solution Approach 2:
Local models are trained in advance on client systems using their respective local datasets before aggregation. This preliminary action enables the system to prepare encrypted parameter sets locally, so that the central server only needs to perform aggregation operations on already-trained models, rather than performing computationally intensive training operations on encrypted data.
2Adaptability or versatility
If federated learning is used to train across multiple systems, then data diversity is improved, but risk of information leakage and auditing complexity worsen
Solution Approach 1:
A central server acts as an intermediary that receives encrypted parameter sets from multiple client systems, aggregates them using secure aggregation protocols, and distributes updated global models back to clients. This intermediary architecture enables diverse data sources to contribute to a global model while the encryption and secure aggregation mechanisms prevent any single party (including the server) from accessing individual client data, thus mitigating information leakage risks.
3Reliability
If personal data is removed for pseudonymization, then privacy is improved, but training quality worsens
Solution Approach 1:
The system changes the state of data from plaintext to encrypted form, allowing personal information to remain in the training datasets while maintaining privacy through encryption. This parameter change enables the use of rich, detailed training data (including personally identifiable information) without sacrificing privacy, as the encryption ensures that even the server cannot access individual data values during aggregation.
Data Source
AI summary
Respective local parameters parametrizing a base function are determined to provide at least one local trained function for each of multiple client systems by training the respective local trained function using machine learning with multiple training datasets on the respective client system, wherein at least some of the training datasets are specific to the respective client system. A respective local plaintext dataset including the local parameters of the respective local trained function is encrypted to generate a respective local encrypted dataset on the respective client system. The local encrypted datasets are transmitted to an aggregating system, and the global encrypted dataset is calculated from the local encrypted datasets using the calculation algorithm.