On-Device Model Query Processing for Private Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing language models on devices face challenges in providing high-quality recommendations while preserving user privacy, as storing the entire model on the device is infeasible due to memory and processing constraints, and traditional privacy protocols are computationally taxing on servers.
Innovation Solution
Implementing a privacy protocol that divides computations between an on-device model and a server, where the on-device model performs a portion of the privacy protocol and the server retrieves raw data, allowing for high-quality recommendations without exposing sensitive information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire language model is stored on the device, then recommendation quality is improved, but memory and processing constraints make it infeasible
Solution Approach 1:
The patent divides the language model into two segments: a compressed version stored on the device and a full version stored on the server. The on-device model handles initial processing and generates queries, while the server-hosted model provides comprehensive information retrieval. This segmentation allows the system to achieve high recommendation quality without requiring the entire model to reside on resource-constrained devices.
2Reliability
If traditional privacy protocols are used, then user privacy is preserved, but server computational overhead increases
Solution Approach 1:
The patent applies preliminary action by having the on-device model perform privacy-preserving transformations and generate refined queries before transmitting them to the server. This preliminary processing reduces the complexity of computations required on the server while maintaining privacy guarantees, as the device prepares data in a privacy-protecting manner before server interaction.
Solution Approach 2:
The on-device model serves as an intermediary between the user and the server. It transforms user inputs and intermediate results through privacy-preserving computations, generating refined queries that the server can process with reduced computational overhead. This intermediary role maintains privacy while optimizing server resource utilization.
3Quantity of substance
If the model resides on the server, then device memory constraints are relieved, but sensitive information may be exposed
Solution Approach 1:
The on-device model acts as a privacy-preserving intermediary that processes sensitive information locally before transmitting data to the server. It performs transformations on user inputs and intermediate results, ensuring that only processed, privacy-protected data leaves the device. This mediator role allows the model to reside on the server while protecting sensitive information through local preprocessing.
Solution Approach 2:
The patent applies parameter changes by transforming data through privacy-preserving computations on the device before transmission. The on-device model modifies the representation of sensitive information through cryptographic or mathematical transformations, changing the parameters of the data so that the server can process useful information without accessing the original sensitive inputs.
Data Source
AI summary
Techniques are disclosed relating to privacy-preserving query processing using on-device models. A device storing a query processing model receives a query. The device sends, based on the query, information requests according to privacy protocols, where the information request is encrypted such that a plaintext version of the given information request is not accessible to the server. The device then receives from the server one or more information responses to the information request that includes response objects generated according to the privacy protocols and are not accessible to the server. The device decrypts, using a cryptographic key, response objects that are received as part of the one or more information responses, generates, using the query processing model and the decrypted response objects, a result for the query. The device then outputs the generated result.