Distributed AI Model Security via Circular Linked List Data Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models are vulnerable to security breaches as they are typically stored in central locations, allowing malicious users to reverse-engineer and manipulate predictions if access is gained.
Innovation Solution
Storing training data across multiple datacenters as a circular linked list and incorporating fake data nodes, where incremental error and bias are used to navigate between datacenters, making it difficult for malicious users to access all data without proper model building.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If the model is stored in a central location, then the model can be easily accessed and retrieved, but the model becomes vulnerable to security breaches and reverse engineering
Solution Approach 1:
The patent divides the training data into multiple data nodes and distributes them across different datacenters. Each data node contains only a portion of the training data and a pointer to the next node, making it difficult for malicious users to access the complete model by compromising a single central location, while still allowing legitimate access through the distributed structure
2Object-affected harmful factors
If the training data is distributed across multiple datacenters, then model security is improved, but the complexity of retrieving and building the model increases
Solution Approach 1:
The patent introduces a model designer as an intermediary component that manages the distributed training data. The model designer retrieves data nodes from multiple datacenters in sequence, builds the model incrementally, and handles the complexity of navigating the distributed structure, thereby shielding end users from the underlying complexity while maintaining security
Solution Approach 2:
The system pre-establishes a circular linked list structure where each data node contains a pointer to the next node. This preliminary organization allows the model designer to systematically retrieve and process data nodes in the correct sequence without requiring complex real-time decision-making about data distribution
3Object-affected harmful factors
If fake training data is appended to the circular linked list, then malicious users cannot determine when all real data has been accessed, but this increases the difficulty of model building
Solution Approach 1:
The patent converts the potential harm of incomplete or incorrect data retrieval into a benefit by using fake data nodes as deliberate obfuscation elements. These fake nodes, interspersed among real training data nodes, prevent malicious users from determining when all real data has been accessed, while the model designer can still build the model correctly by following the structured sequence and using error/bias calculations to identify valid data nodes
Data Source
AI summary
A model designer improves the security of a machine learning model in certain embodiments. Instead of storing the model in a central location, the training data used to build and train the model is stored across several different databases and/or datacenters. The training data is divided into portions and stored as a circular linked list across these databases and/or datacenters. The model designer retrieves the training data and incrementally builds and trains the model using the training data. The incremental error and bias of the model is used to locate training data between datacenters. Additionally, fake training data is appended to the circular linked list and the model designer tracks how much training data is used before hitting fake training data.