Secure Federated Machine Learning With Enclave-Based Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning systems require data exchange between entities, raising privacy concerns and potential security risks, as entities may not wish to share their data due to privacy and security issues.
Innovation Solution
A secure platform computer with a secure memory space enables entities to collaborate on projects by storing and processing data within an enclave, ensuring data privacy through cryptographic hashing and access control, allowing the generation and utilization of a machine-learning model without exchanging raw data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple entities exchange data to train a machine-learning model, then the model can be generated collaboratively, but privacy concerns and security risks arise
Solution Approach 1:
The patent introduces a trusted third-party platform that acts as an intermediary to facilitate collaborative machine-learning model training. This platform enables multiple entities to contribute their data and computational resources without directly exchanging raw data between themselves. The intermediary platform processes encrypted data and coordinates the federated learning process, thereby achieving collaborative model generation while protecting entity privacy and reducing security risks associated with direct data exchange
2Measurement precision
If entities share their data for machine-learning projects, then the model quality improves, but data privacy is compromised
Solution Approach 1:
The patent segments the machine-learning training process into distributed components that operate across multiple entities. Each entity maintains local data and trains local model instances independently, then shares only model parameters or gradients with the central platform. This segmentation allows the system to aggregate knowledge from multiple sources to improve model quality while ensuring that raw private data remains localized and never leaves the owning entity's secure environment
Solution Approach 2:
The patent employs copying mechanisms where entities create and share copies of their local model parameters, gradients, or encrypted data representations rather than sharing the actual raw data. The central platform aggregates these copies to train the global model. This approach enables model quality improvement through collaborative learning while the original private data remains untouched and confidential at each entity's location
3Reliability
If a secure platform is introduced to protect data privacy, then security is improved, but system complexity increases
Solution Approach 1:
The patent designs the secure platform to perform multiple functions within a single integrated system. The platform simultaneously provides cryptographic key management, secure communication channels, coordinated training orchestration, model aggregation, and result distribution. By consolidating these diverse security and computational functions into one multi-functional platform, the system achieves high data security and reliability while avoiding the need for multiple separate complex security systems that would otherwise be required
Data Source
AI summary
Embodiments of the invention are directed to systems, methods, and devices for securely performing federated tasks (e.g., the generation and utilizing of machine-learning models). A secure platform computer may operate a secure memory space. Entities participating in a federated project may transmit respective portions of project data defining the federated project. Each entity may provide their respective (encrypted) data sets for the project that in turn can be used to generate a machine-learning model in accordance with the project data. The machine-learning model may be stored in the secure memory space and accessed through an interface provided by the secure platform computer Utilizing the techniques discussed herein, a machine-learning models may be generated and access to these models may be restricted while protect each participant's data set from being exposed to the other project participants.