Secure ML Workflow Automation Using Isolated Virtual Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning applications face challenges in coordinating repetitive tasks across teams and ensuring data security, particularly when handling sensitive information like personally identifiable information (PII), which requires stringent security measures to prevent data leakage and unauthorized access.
Innovation Solution
A secure machine learning automation service (SMLAS) is implemented within a provider network, utilizing isolated virtual networks and dynamically provisioned computing services to automate machine learning workflows, ensuring that sensitive data is not accessible from untrusted platforms, and performing sanitization operations to remove sensitive information from model outputs before deployment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If real-world sensitive data is used for training machine learning models, then model quality is improved, but data security and privacy risks increase
Solution Approach 1:
The system segments the machine learning workflow into distinct isolated environments: a development environment for model creation and an isolated environment for training with sensitive data. This segmentation allows model quality improvement using real data while maintaining security through environmental separation. The isolated environment is dynamically created, used for training, and then destroyed, preventing data leakage.
Solution Approach 2:
The system introduces an intermediary mechanism (the isolated environment with dynamic resource provisioning) between the sensitive data and the machine learning models. This intermediary ensures that sensitive data never persists outside the controlled isolated environment, while still allowing models to access and learn from the data during the training process, thus resolving the conflict between data quality and security.
2Adaptability or versatility
If multiple teams coordinate repetitive machine learning tasks, then comprehensive model development is achieved, but coordination effort and time increase
Solution Approach 1:
The system implements self-service automation where the isolated environment automatically provisions computing resources, executes training workflows, and destroys itself after completion. This eliminates the need for manual coordination between teams for repetitive tasks like data preparation, model training, and evaluation. The automated workflow engine manages the entire process, reducing coordination overhead while maintaining comprehensive model development capabilities.
Solution Approach 2:
The isolated environment is designed as a universal platform that can handle multiple machine learning tasks (training, evaluation, hyperparameter tuning) and support various data types and model architectures. This multi-functionality allows a single automated system to replace multiple specialized team workflows, reducing coordination complexity while maintaining versatility in model development.
3Ease of manufacture
If cloud-based distributed resources are used for machine learning, then infrastructure cost is reduced, but data security control becomes more challenging
Solution Approach 1:
The system employs dynamic resource provisioning where computing resources are created on-demand within the isolated environment, used for the specific training task, and then automatically destroyed. This dynamic approach maintains the cost benefits of cloud-based distributed resources while ensuring data security through the temporary nature of resource existence. Sensitive data never persists in the cloud infrastructure, as both the data and computing resources are ephemeral and confined to the isolated environment.
Data Source
AI summary
An indication of training artifacts for a machine learning model to be trained with an input data set having an access restriction is obtained. A representation of a software execution environment containing the artifacts is deployed to a computing platform within an isolated resource group which satisfies the access restriction. A trained version of the machine learning model is generated at the computing platform, and transferred outside the isolated resource group.