Annotation-Based Data Masking for Secure LLM Prompt Input
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems that utilize generative machine learned models (LLMs) face issues when inputting sensitive data, which can compromise user experience and security by exposing confidential information.
Innovation Solution
A data masking component that uses data annotations to identify and mask sensitive data before inputting it into LLMs, employing templates with slots, metadata classification, and policies to determine masking criteria.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sensitive data is input into generative machine learned models, then the model can process and generate responses using the data, but user experience and security deteriorate due to exposure of confidential information
Solution Approach 1:
The system performs preliminary actions by identifying and masking sensitive data in the input prompt before it is processed by the generative machine learned model. The data masking component scans the input data, detects sensitive information based on predefined criteria, and replaces it with masked versions, ensuring that the model receives sanitized input that maintains processing capability while protecting confidentiality.
2Reliability
If data masking is implemented using annotations and policies, then data security and user experience are improved, but system complexity increases due to additional processing components
Solution Approach 1:
The data masking component serves as an intermediary between the user input and the generative machine learned model. It acts as a mediator that receives the original input, applies masking transformations based on annotations and policies, and forwards the sanitized version to the model. This intermediary approach protects the model from sensitive data while maintaining the overall system architecture.
Solution Approach 2:
The system segments the data processing function into distinct components: the original input processing, the data masking component with its annotation and policy evaluation logic, and the generative model. This segmentation allows each component to specialize in its function, with the masking component handling security concerns independently from the model's core processing tasks.
3Reliability
If all data is masked to ensure security, then data confidentiality is maintained, but information loss occurs reducing model effectiveness
Solution Approach 1:
The system applies local quality by selectively masking only the sensitive portions of the input data while leaving non-sensitive information unchanged. The data masking component evaluates each data element against predefined criteria and annotations, applying masking transformations only where necessary to protect confidentiality, thereby preserving the utility and effectiveness of the overall input for model processing.
Data Source
AI summary
Techniques for masking data based on annotations are discussed herein. A system may receive a request to perform an action and leverage an LLM to assist in performing the requested action. When generating the input data to input to the LLM, the system can use a template to organize the input data. The template may include static data and/or slot(s) which can include a reference to data to input into such slots. The system may retrieve data to input to the slot based on the reference, retrieve annotations that define a classification of the data, and receive a policy that defines which types of data classifications are to be masked. Based on the data classification and the policy, the system can determine whether to mask the data. The system can generate the input data using the template, the data, and/or the mask(s) and input such data into the LLM.