Multi-tenant Machine Learning Model Storage with Encryption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning platforms face challenges in efficiently training and deploying models while ensuring security and resource optimization, particularly in handling multiple customers' sensitive data and managing model versions effectively.
Innovation Solution
A machine learning platform with a multi-tenant model store and scheduler that securely trains, stores, and deploys models, using versioning and encryption to isolate customer data and dynamically update models, ensuring efficient resource use and security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple customers' data is processed on shared training and prediction servers, then resource utilization and productivity are improved, but security and data isolation are compromised
Solution Approach 1:
The patent segments customer data and models into isolated containers on shared servers. Each customer's data, trained models, and predictions are separated using unique identifiers (e.g., customer IDs, encryption keys) that prevent cross-contamination while allowing multiple customers to coexist on the same physical infrastructure. This enables secure multi-tenant architecture where resource sharing does not compromise security.
Solution Approach 2:
The patent introduces intermediary mechanisms such as encryption layers, authentication protocols, and managed storage systems that mediate between multiple customers and the shared server infrastructure. These intermediaries ensure that each customer's data remains protected while allowing efficient resource sharing, resolving the contradiction between security and resource utilization.
2Speed
If trained models are stored on prediction servers with customer data, then access speed is improved, but security risks and data exposure increase
Solution Approach 1:
The patent extracts trained models from customer data storage locations and places them in a separate, secure model repository. Models are stored independently of raw customer data, but can be quickly retrieved and applied to predictions. This separation reduces data exposure risk while maintaining fast access through optimized model loading and caching mechanisms.
Solution Approach 2:
The patent implements preliminary actions by pre-training models securely and storing them in ready-to-use format in a dedicated model store. Models are prepared and validated before being deployed to prediction servers, ensuring they are secure and optimized for fast retrieval. This preliminary preparation enables rapid prediction without requiring access to raw customer data during the prediction process.
3Productivity
If training data is retained on training servers for future use, then productivity is improved, but security risks and resource consumption increase
Solution Approach 1:
The patent implements a model lifecycle management system where trained models are extracted from training servers and stored in a dedicated model repository. After training completes, the training data and temporary processing resources on training servers can be discarded or returned to the pool, reducing resource consumption. The trained models are recovered and stored in an optimized format for future prediction use, maintaining productivity without sustained resource burden.
Data Source
AI summary
A request to perform a prediction using a machine learning model of a specific entity is received. A specific security key for the machine learning model of the specific entity is received. At least a portion of the machine learning model is obtained from a multi-tenant machine learning model storage. The machine learning model is unlocked using the specific security key and the requested prediction is performed. A result of the prediction is provided from a prediction server.