Privacy-Preserving Search Suggestions Using LLM Phrase Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional search engines expose sensitive information and user interest by suggesting search queries based on past search data, leading to potential privacy breaches and community exposure.
Innovation Solution
A system and method using Large Language Models (LLMs) to generate context-relevant search suggestions without relying on user-provided information, employing word embeddings, deduplication, and evaluation based on search results to provide privacy-preserving search suggestions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If search engines use past search data to provide search suggestions, then search suggestions become more relevant and helpful, but user privacy is compromised and sensitive information is exposed
Solution Approach 1:
The patent introduces an intermediary mechanism using anonymized query logs and aggregated data processing. Instead of directly using user search history, the system processes queries through an intermediate layer that removes identifiable information while preserving search pattern insights, thus enabling relevant suggestions without direct access to user privacy data
Solution Approach 2:
The system creates a simplified representation or copy of search data that lacks sensitive user information. By working with anonymized and aggregated versions of query data rather than raw user searches, the system can derive useful patterns for suggestions while the original user data remains protected and inaccessible
2Measurement precision
If search engines process user search history to generate suggestions, then search accuracy improves, but interest between different user communities is exposed
Solution Approach 1:
The patent segments search data by treating different user communities as separate, isolated data streams. By processing queries within discrete user contexts and using personalized anonymized profiles, the system maintains accurate community-specific suggestions while preventing cross-community information leakage and interest exposure
Solution Approach 2:
The system applies local quality processing by customizing search suggestion generation for each user's specific context and community while using the same overall anonymized framework. This allows high-accuracy community-relevant suggestions to be generated locally for each user without exposing broader community interests to other communities
3Adaptability or versatility
If search engines rely on past search information, then search suggestions become more personalized, but user privacy and sensitivity are compromised
Solution Approach 1:
The patent uses anonymized query logs as an intermediary that mediates between user search behavior and suggestion generation. This intermediary layer enables personalized suggestions by capturing individual search patterns while removing all directly identifiable sensitive information, thus maintaining personalization without compromising user privacy
Data Source
AI summary
A system and method for providing privacy-preserving search suggestions is disclosed. The system receives a plurality of documents having text content and generates at least one first word embedding for each document. The system further generates a list of first search phrases for each document using Large Language Models (LLMs), and generates at least one second word embedding for each first search phrase. Further, each first word embedding is compared to the corresponding second word embedding to rank the first search phrases based on similarity to the documents. The system is configured to deduplicate one or more ranked search phrases having a rank lower than a first predefined rank, and execute remaining ranked search phrases after deduplication in a search engine to evaluate search results and determine final search phrases from the remaining ranked search phrases based on the search results.