Untrusted ML Model Sandboxing for Secure Distributed Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data analytic platforms lack a framework to efficiently support multiple programming languages without search latency, and there is no mechanism to deploy ML models and streaming algorithms within search workflows independently of programming language or origin, while ensuring secure execution of untrusted ML models in distributed environments.
Innovation Solution
A data intake and query system that uses a late-binding schema and flexible extraction rules to process and index machine data, enabling secure deployment of ML models and streaming algorithms across diverse data systems, and ensures secure execution of untrusted code in distributed environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If untrusted ML models are executed in distributed runtime environments to enable resource sharing and computing efficiencies, then productivity and resource utilization are improved, but security and trust reliability deteriorate due to potential malicious code access
Solution Approach 1:
The patent introduces a sandboxed execution environment as an intermediary layer between the untrusted ML model code and the distributed runtime system. This sandbox acts as a mediator that allows execution of untrusted code while preventing direct access to system resources, data, and other tenants, thus resolving the contradiction between enabling productivity and maintaining security.
Solution Approach 2:
The execution environment is segmented into isolated sandboxes, each containing a specific untrusted ML model. This segmentation ensures that malicious code in one sandbox cannot access or harm other sandboxes or the underlying system, allowing secure concurrent execution of multiple untrusted models in the distributed runtime environment.
2Adaptability or versatility
If data is stored in minimal processed form to enable flexible later analysis, then adaptability and data flexibility are improved, but search and analysis performance deteriorates due to the vast amount of data requiring processing
Solution Approach 1:
The patent segments the data processing into two distinct phases: (1) an indexing phase where data is processed and transformed into a search-optimized format, and (2) a query execution phase where only the necessary segmented data is retrieved and processed. This segmentation allows the system to maintain data flexibility while improving search performance by avoiding processing the entire dataset at query time.
Solution Approach 2:
The system performs preliminary processing and indexing of data before actual queries are executed. During the indexing phase, data is pre-processed, transformed, and organized in a way that optimizes for future search operations. This preliminary action enables fast query execution without requiring reprocessing of the entire dataset, thus resolving the contradiction between data flexibility and search performance.
3Adaptability or versatility
If multiple programming languages are supported in the data analytic platform to increase versatility, then adaptability is improved, but search latency increases due to programming language translations
Solution Approach 1:
The patent introduces a common query language as an intermediary that all programming languages translate into during the indexing phase. This common language serves as the internal representation for data queries, eliminating the need for real-time translation during search execution. The intermediary approach allows multi-language support while minimizing latency by using a standardized query format for all operations.
Data Source
AI summary
A computer-implemented method is disclosed that includes operations of receiving a request to store a representation of a machine learning (ML) model in a non-transitory computer-readable medium, validating the representation of the ML model, and storing the representation of the ML model. The representation of the ML model may be a computation graph, and further, the computation graph may be a format for representing ML models through a defined common set of operators and a common file format. Validating the representation of the ML model may include verifying an input schema of the representation of the ML model is compatible with an input schema of the ML model. The method may include further operations of deserializing the representation of the ML model, wherein a successful deserialization indicates that the representation of the ML model is valid and secure.