Split Inference Model Deployment to Limit Sensitive Data Reconstruction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing inference models are vulnerable to unauthorized access, allowing sensitive information to be reconstructed, compromising data security and compliance with varying privacy regulations.

Innovation Solution

Implementing an input data attack resistant inference model with a shared body portion deployed at a trusted location and a prediction head portion at a potentially untrustworthy location, trained to impart reconstruction resistance based on sensitivity levels, reducing the likelihood of sensitive information being compromised.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If an inference model is deployed at a potentially untrustworthy location to provide computer-implemented services, then service availability and accessibility are improved, but the risk of sensitive information being reconstructed increases

Engineering Contradiction:

Improveservice availabilityVSAvoidreconstructability of sensitive information

Core Design Contradiction:

ProductivityVSObject-affected harmful factors

Solution Approach 1:

The inference model is divided into two distinct portions: a shared body portion deployed at a trusted location and a prediction head portion deployed at the potentially untrustworthy location. This segmentation allows the system to maintain service availability at the untrustworthy location while protecting sensitive information by keeping the vulnerable components separated from the trusted data source.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The shared body portion acts as an intermediary between the input data and the prediction head portion. It processes input data at the trusted location and transmits only necessary intermediate representations to the prediction head, preventing direct access to sensitive information while enabling service functionality at the untrustworthy location.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If the inference model is trained to impart reconstruction resistance to input features, then data security is improved, but the model complexity and training requirements increase

Engineering Contradiction:

Improvedata securityVSAvoidmodel complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

Reconstruction resistance is imparted selectively to specific portions of the model architecture. The shared body portion is trained with reconstruction resistance properties for sensitive input features, while the prediction head portion maintains standard predictive functionality. This localized application of security properties reduces overall model complexity compared to making the entire model resistant to reconstruction.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The shared body portion is trained in advance to impart reconstruction resistance to input features before the prediction head portion is trained. This preliminary training establishes protective properties that prevent sensitive information from being reconstructed, simplifying the subsequent training of the prediction head and reducing overall system complexity.

Inventive Principle:

Principle #10Preliminary action

3Object-affected harmful factors

If the shared body portion is kept at a trusted location and only the prediction head is deployed remotely, then reconstructability of sensitive information is reduced, but communication overhead and latency increase

Engineering Contradiction:

Improvereconstructability of sensitive informationVSAvoidcommunication latency

Core Design Contradiction:

Object-affected harmful factorsVSLoss of time

Solution Approach 1:

The prediction head portion, which contains the components most vulnerable to reconstruction attacks, is extracted and deployed remotely at the untrustworthy location. Only the essential shared body portion remains at the trusted location, minimizing the amount of data that needs to be transmitted back and forth while maintaining security.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The shared body portion performs preliminary processing of input data at the trusted location, extracting and transmitting only the necessary intermediate representations to the remote prediction head. This preliminary action reduces the volume of data communication required, thereby minimizing communication overhead and latency while maintaining security.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12579261B2Managing inference models in view of reconstructability of sensitive information

Publication Date: 2026.03.17 DELL PROD LP

AI summary

Methods, systems, and devices for providing computer-implemented services are disclosed. To provide the computer-implemented services, inference models may be deployed to locations. Prior to deploying an inference model to a location, it may be determined whether the location is trustworthy. If the location is determined to not be trustworthy, an input data attack resistant inference model may be selected and deployed. The input data attack resistant inference model may be trained to impart reconstruction resistance to sensitive information during inference generation based on a schema for weighting the sensitive information for reconstruction resistance. The training process may decrease a likelihood of the inferences generated by the input data attack resistant inference model being usable to reconstruct the sensitive information.