Protected Server Environment for Secure Machine Learning Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems face challenges in protecting sensitive training data, particularly when stringent regulations like HIPAA require secure handling of medical claims data, and there is a need to ensure that personal information is not inadvertently exposed during training and validation processes.
Innovation Solution
A protected environment is established where training data is stored and processed, using de-identification and encryption techniques to generate a machine learning system that satisfies specific criteria for data protection, ensuring that only a trained and validated system is released, which cannot memorize inputs and maintains data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is made accessible for machine learning training and validation, then the machine learning system can be trained and validated effectively, but sensitive personal information may be exposed or memorized by the system
Solution Approach 1:
A trusted third-party server acts as an intermediary between data holders and machine learning system developers. The server hosts the training data in a protected environment, allowing the machine learning system to access and learn from the data without any party obtaining direct access to the sensitive information. This mediator approach enables effective training while maintaining data privacy.
Solution Approach 2:
Instead of sharing actual sensitive data, the system creates and uses copies or representations of the data in a controlled environment. The machine learning system trains on replicated data structures that preserve learning capabilities while eliminating direct exposure to sensitive personal information through de-identification and encryption techniques.
2Object-affected harmful factors
If data protection measures like de-identification and encryption are applied to training data, then data privacy is protected, but the complexity of the training process increases
Solution Approach 1:
De-identification and encryption measures are applied to the training data before it is uploaded to the protected environment. This preliminary processing ensures data protection is built into the foundation of the training process, rather than adding complexity during active training. The data is prepared in advance with protective measures already in place.
Solution Approach 2:
The protected environment automatically manages encryption keys, de-identification mappings, and security protocols without requiring manual intervention during training. The system self-manages the complexity of data protection measures, allowing users to benefit from security without directly handling the computational overhead.
3Object-affected harmful factors
If a machine learning system is trained on protected data without direct access, then data security is maintained, but validating the system's performance becomes difficult
Solution Approach 1:
The protected environment implements feedback mechanisms where the machine learning system's predictions and performance metrics are returned to validators without exposing the underlying sensitive data. This allows validation of system effectiveness through performance feedback while maintaining data security through the intermediary architecture.
Data Source
AI summary
A computer-implemented method comprising: storing, using a server computer executing within a protected environment, a plurality of media items, each of the media items corresponding to one of a plurality of different status values; receiving, from a requesting computing device that is outside the protected environment, a request to send certain media items outside the protected environment to a client computing device; computing, using a plurality of machine learning systems executed by the server computer, each of the machine learning systems having been trained with one of the plurality of status values as an output, a plurality of likelihood values associated with a particular status value for the client computing device, each of the machine learning systems having been trained at least in part using attribute values associated with health data records as inputs, and an existence or a non-existence of a one of the plurality of different status values as outputs, the server computer storing first data comprising a plurality of attribute values for a plurality of the health data records and second data indicating, for each health data record of the plurality of health data records, whether the health data record has the status value, the server computer being configured to train the particular machine learning system in the protected environment only if the first data and the second data satisfy a first criterion and being configured to send the particular machine learning system to the requesting computing device only if the particular machine learning system satisfies a second criterion; identifying a particular status value, among the plurality of status values, having a highest likelihood value; selecting a specific set of media items at least partly based on the identified particular status value having the highest likelihood value, in a number indicated by the request to send certain media items outside the protected environment to the client computing device; and sending, from the server computer to the client computing device, the specific set of media items that have been selected.