k-Aggregation Anonymization for Cross-Regional ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data anonymization techniques fail to reliably meet the privacy requirements of singling-out, linkability, and inference prevention, limiting the ability of machine learning models to cross regional boundaries and retain data for extended periods.
Innovation Solution
A novel k-aggregation data anonymization process that randomly selects and aggregates data samples from multiple enterprises and users, repeating this process to generate anonymized datasets classified as system metadata, which can be used to train machine learning models that provide feed ranking and content recommendation services while ensuring privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing data anonymization techniques are used, then data privacy requirements are partially met, but the techniques fail to reliably prevent singling-out, linkability, and inference
Solution Approach 1:
The patent combines multiple data samples from different users and enterprises into aggregated data samples. By merging k data samples through random selection and aggregation operations, the system creates anonymized data that prevents singling-out individual users while maintaining statistical properties useful for machine learning training.
Solution Approach 2:
The patent introduces an intermediary anonymization process between raw customer content data and machine learning training data. This intermediary k-aggregation process transforms identifiable customer data into anonymized system metadata that serves as a mediator, enabling ML training while preventing direct access to individual user information.
2Manufacturing precision
If customer content data is used for machine learning training, then model quality improves, but privacy requirements for singling-out, linkability, and inference prevention are not met
Solution Approach 1:
The patent changes the parameters of data representation by transforming raw customer content data into aggregated statistical representations. By changing from individual user data parameters to aggregated group parameters (through k-anonymity and differential privacy mechanisms), the system maintains data utility for ML training while satisfying privacy requirements.
3Reliability
If data is anonymized using traditional techniques, then some privacy protection is achieved, but the anonymized data cannot be retained for extended periods or crossed regional boundaries
Solution Approach 1:
The patent applies preliminary anonymization actions (k-aggregation and differential privacy) to data before storage and cross-regional transfer. By performing anonymization in advance, the system enables long-term retention and cross-boundary usage of data while maintaining privacy protection throughout the data lifecycle.
4Reliability
If k-aggregation is performed with larger k values, then privacy protection improves, but data utility for training machine learning models decreases
Solution Approach 1:
The patent implements dynamic adjustment of aggregation parameters and privacy mechanisms. By making the anonymization process adaptive rather than static, the system can optimize the balance between privacy protection (k value) and data utility based on specific training requirements and privacy risk assessments.
Data Source
AI summary
A method for transforming customer content data to anonymized system metadata includes causing execution of an enterprise application on remote computing systems operated by users associated with multiple enterprises and logging customer content data including data samples corresponding to the users' interactions with the enterprise application. The method includes performing k-aggregation of the data samples by: (a) randomly selecting an enterprise; (b) randomly selecting a user associated with the enterprise; (c) randomly selecting a data sample of the user; (d) repeating (a), (b), and (c) k times, where (a) and (b) are performed without replacement; and aggregating the randomly-requested data samples by position. The method includes repeating the k-aggregation N times with replacement to generate N aggregated data samples, concatenating such data samples to generate an anonymized dataset, training a machine learning model using the anonymized dataset, and deploying the trained machine learning model via the enterprise application.