Secure Multi-User Machine Learning on Cloud Data Platform
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing cloud-based database systems face challenges in securely sharing data between users to create machine learning models due to technical complexity, scaling issues, energy inefficiencies, and privacy concerns, with previous approaches being complex, costly, and slow to implement.
Innovation Solution
A secure machine learning system that allows users to share encrypted data and functions, enabling the generation of machine learning models without revealing underlying data, using a network-based data warehouse system with a compute service manager and execution platform to manage and process data securely across multiple users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If data is shared between users to create machine learning models, then model accuracy and functionality are improved, but security and privacy risks increase
Solution Approach 1:
The patent introduces a trusted intermediary system (the database system itself) that enables secure multi-party machine learning. The intermediary provides encrypted data sharing capabilities, allowing users to train models on combined data without exposing individual datasets. The system includes cryptographic primitives and access control mechanisms that mediate between data sharing needs and security requirements.
Solution Approach 2:
The patent employs encrypted copies of data for machine learning operations. Instead of sharing original data, the system creates and manipulates encrypted representations of data throughout the machine learning process. The data remains encrypted during training, and only decrypted results are returned to users, enabling model functionality while preserving data security.
2Object-affected harmful factors
If complex encryption and security measures are implemented, then data privacy is improved, but system complexity and implementation difficulty increase
Solution Approach 1:
The patent designs a universal security framework that handles multiple privacy requirements through a single integrated system. The same cryptographic infrastructure supports various access patterns, encryption modes, and data sharing scenarios. This multi-functional approach reduces overall system complexity compared to implementing separate security solutions for each use case.
Solution Approach 2:
The system dynamically adjusts security parameters based on access requests and user roles. Instead of using fixed encryption schemes, the patent implements flexible parameter changes in encryption strength, data access permissions, and cryptographic operations based on the specific machine learning task and user requirements, simplifying implementation while maintaining privacy.
3Object-affected harmful factors
If data is encrypted and shared securely, then privacy protection is improved, but computational overhead and processing time increase
Solution Approach 1:
The patent performs cryptographic operations in advance during data loading and preprocessing stages. Encryption keys are established and data is encrypted before the actual machine learning training begins. This preliminary action reduces the computational burden during training, as the system can work with pre-encrypted data without repeated cryptographic operations, thereby reducing processing time while maintaining privacy.
Data Source
AI summary
A system for providing access to a database management system (DBMS) to a first user of a cloud data platform, the DBMS being generated by a second user. A machine learning model for training on a training dataset is included in the DBMS. The training dataset includes a first training dataset that is encrypted in the DBMS and a second training dataset that includes non-overlapping features with the first training dataset. A request, from the second user, to train the machine learning model on the first and second training datasets is identified. A trained machine learning model is generated by training the machine learning model on a joined dataset according to the request. One or more outputs from the trained machine learning model are generated by applying the trained machine learning model on new data.