Multipath Inference Models for Untrusted-Location Data Protection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unauthorized entities may attempt to reconstruct sensitive input data used to generate inferences by inference models, compromising data security and privacy due to vulnerabilities in hardware resources, network security, and differing data privacy regulations.
Innovation Solution
Deploy a multipath inference model with a shared body portion at a trusted location and a prediction head portion at an untrusted location, trained to obfuscate input data during inference generation, using a modified split training process to reduce the reconstructive power of the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If an inference model is deployed at an untrusted location to provide computer-implemented services, then service accessibility and operational flexibility are improved, but the risk of unauthorized reconstruction of sensitive input data increases
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 an untrusted location. This segmentation allows the system to maintain service accessibility at untrusted locations while protecting sensitive input data by keeping the vulnerable reconstruction pathway at the trusted location only.
Solution Approach 2:
The shared body portion acts as an intermediary between the input data at the trusted location and the prediction head at the untrusted location. It processes the input data and generates intermediate outputs that are sufficient for the prediction head to generate accurate predictions, yet insufficient for unauthorized entities to reconstruct the original input data.
2Reliability
If the shared body portion is frozen during training to reduce reconstructive power, then data security is improved, but the training time and computational resources increase
Solution Approach 1:
The shared body portion is frozen before the final training stage, performing the preliminary action of establishing a secure, immutable foundation. This allows the prediction head to be trained independently on untrusted data without risking back-propagation that could enable reconstruction attacks, thereby securing the system architecture early in the training process.
Solution Approach 2:
The training process dynamically adjusts the state of the model components: the shared body portion transitions from a trainable state to a frozen state, while the prediction head remains trainable throughout. This dynamic approach optimizes both security (through freezing) and training efficiency (by allowing parallel training of the prediction head).
Data Source
AI summary
Methods, systems, and devices for providing computer-implemented services are disclosed. To provide the computer-implemented services, inference models used by data processing systems may be deployed to locations to generate inferences. 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 to perform inference generation. The input data attack resistant inference model may be trained to decrease a likelihood of the inferences generated by the input data attack resistant inference model being usable to reconstruct input data used to generate the inferences. By doing so, sensitive input data may be obfuscated during inference generation.